improved provider picker ui on resource

add streaming function to capability
add some parallelization to ai gateway pipeline
2026-08-08 13:38:33 +02:00 · 2026-08-07 16:13:51 -04:00 · 2026-08-07 14:36:15 -04:00 · 2026-08-07 14:36:11 -04:00 · 2026-08-07 13:54:32 -04:00 · 2026-08-07 12:25:29 -04:00
174 changed files with 24100 additions and 868 deletions
@@ -1,5 +1,5 @@
-# FROM node:24-slim AS base
-FROM public.ecr.aws/docker/library/node:24-slim AS base
+# FROM node:24.18.1-slim AS base
+FROM public.ecr.aws/docker/library/node:24.18.1-slim AS base

 WORKDIR /app

@@ -32,8 +32,8 @@ FROM base AS builder

 RUN npm ci --omit=dev

-# FROM node:24-slim AS runner
-FROM public.ecr.aws/docker/library/node:24-slim AS runner
+# FROM node:24.18.1-slim AS runner
+FROM public.ecr.aws/docker/library/node:24.18.1-slim AS runner

 WORKDIR /app

@@ -1,4 +1,4 @@
-FROM node:24-alpine
+FROM node:24.18.1-alpine

 WORKDIR /app

@@ -1,5 +1,5 @@
 import { CommandModule } from "yargs";
-import { db, idpOidcConfig, licenseKey, certificates, eventStreamingDestinations, alertWebhookActions } from "@server/db";
+import { db, idpOidcConfig, licenseKey, certificates, eventStreamingDestinations, alertWebhookActions, aiProviders } from "@server/db";
 import { encrypt, decrypt } from "@server/lib/crypto";
 import { configFilePath1, configFilePath2 } from "@server/lib/consts";
 import { eq } from "drizzle-orm";
@@ -132,12 +132,14 @@ export const rotateServerSecret: CommandModule<
            const certs = await db.select().from(certificates);
            const streamingDestinations = await db.select().from(eventStreamingDestinations);
            const webhookActions = await db.select().from(alertWebhookActions);
+            const providers = await db.select().from(aiProviders);

            console.log(`Found ${idpConfigs.length} OIDC IdP configuration(s)`);
            console.log(`Found ${licenseKeys.length} license key(s)`);
            console.log(`Found ${certs.length} certificate(s)`);
            console.log(`Found ${streamingDestinations.length} event streaming destination(s)`);
            console.log(`Found ${webhookActions.length} alert webhook action(s)`);
+            console.log(`Found ${providers.length} AI provider(s)`);

            // Prepare all decrypted and re-encrypted values
            console.log("\nDecrypting and re-encrypting values...");
@@ -171,11 +173,18 @@ export const rotateServerSecret: CommandModule<
                encryptedConfig: string;
            };

+            type AiProviderUpdate = {
+                providerId: number;
+                encryptedApiKey: string | null;
+                encryptedHeaders: string | null;
+            };
+
            const idpUpdates: IdpUpdate[] = [];
            const licenseKeyUpdates: LicenseKeyUpdate[] = [];
            const certUpdates: CertUpdate[] = [];
            const streamingDestinationUpdates: StreamingDestinationUpdate[] = [];
            const webhookActionUpdates: WebhookActionUpdate[] = [];
+            const aiProviderUpdates: AiProviderUpdate[] = [];

            // Process idpOidcConfig entries
            for (const idpConfig of idpConfigs) {
@@ -306,6 +315,37 @@ export const rotateServerSecret: CommandModule<
                }
            }

+            // Process aiProviders entries (apiKey + headers)
+            for (const provider of providers) {
+                try {
+                    if (!provider.apiKey && !provider.headers) {
+                        continue;
+                    }
+
+                    const encryptedApiKey = provider.apiKey
+                        ? encrypt(decrypt(provider.apiKey, oldSecret), newSecret)
+                        : null;
+                    const encryptedHeaders = provider.headers
+                        ? encrypt(
+                              decrypt(provider.headers, oldSecret),
+                              newSecret
+                          )
+                        : null;
+
+                    aiProviderUpdates.push({
+                        providerId: provider.providerId,
+                        encryptedApiKey,
+                        encryptedHeaders
+                    });
+                } catch (error) {
+                    console.error(
+                        `Error processing AI provider ${provider.providerId}:`,
+                        error
+                    );
+                    throw error;
+                }
+            }
+
            // Perform all database updates in a single transaction
            console.log("\nUpdating database in transaction...");
            await db.transaction(async (trx) => {
@@ -376,6 +416,17 @@ export const rotateServerSecret: CommandModule<
                            )
                        );
                }
+
+                // Update AI provider entries
+                for (const update of aiProviderUpdates) {
+                    await trx
+                        .update(aiProviders)
+                        .set({
+                            apiKey: update.encryptedApiKey,
+                            headers: update.encryptedHeaders
+                        })
+                        .where(eq(aiProviders.providerId, update.providerId));
+                }
            });

            console.log(`Rotated ${idpUpdates.length} OIDC IdP configuration(s)`);
@@ -383,6 +434,7 @@ export const rotateServerSecret: CommandModule<
            console.log(`Rotated ${certUpdates.length} certificate(s)`);
            console.log(`Rotated ${streamingDestinationUpdates.length} event streaming destination(s)`);
            console.log(`Rotated ${webhookActionUpdates.length} alert webhook action(s)`);
+            console.log(`Rotated ${aiProviderUpdates.length} AI provider(s)`);

            // Update config file with new secret
            console.log("\nUpdating config file...");
@@ -402,6 +454,7 @@ export const rotateServerSecret: CommandModule<
            console.log(`  - Certificates: ${certUpdates.length}`);
            console.log(`  - Event streaming destinations: ${streamingDestinationUpdates.length}`);
            console.log(`  - Alert webhook actions: ${webhookActionUpdates.length}`);
+            console.log(`  - AI providers: ${aiProviderUpdates.length}`);
            console.log(
                `\n  IMPORTANT: Restart the server for the new secret to take effect.`
            );
@@ -1424,6 +1424,16 @@
    "actionDeleteSite": "Delete Site",
    "actionGetSite": "Get Site",
    "actionListSites": "List Sites",
+    "actionCreateAiProvider": "Create AI Provider",
+    "actionDeleteAiProvider": "Delete AI Provider",
+    "actionGetAiProvider": "Get AI Provider",
+    "actionListAiProviders": "List AI Providers",
+    "actionUpdateAiProvider": "Update AI Provider",
+    "actionCreateAiModel": "Create AI Model",
+    "actionDeleteAiModel": "Delete AI Model",
+    "actionGetAiModel": "Get AI Model",
+    "actionListAiModels": "List AI Models",
+    "actionUpdateAiModel": "Update AI Model",
    "actionApplyBlueprint": "Apply Blueprint",
    "actionListBlueprints": "List Blueprints",
    "actionGetBlueprint": "Get Blueprint",
@@ -1622,6 +1632,171 @@
    "sidebarInvitations": "Invitations",
    "sidebarRoles": "Roles",
    "sidebarShareableLinks": "Shareable Links",
+    "sidebarAiGateway": "AI Gateway",
+    "sidebarAiProviders": "Providers",
+    "commandAiProviders": "AI Providers",
+    "aiProvidersTitle": "AI Providers",
+    "aiProvidersDescription": "Connect model providers for AI workloads in this organization",
+    "aiProvidersAdd": "Add Provider",
+    "aiProvidersSearch": "Search providers...",
+    "aiProvidersEmpty": "No AI providers yet",
+    "aiProviderCreate": "Create AI Provider",
+    "aiProviderCreateDescription": "Add a model provider for this organization",
+    "aiProviderSeeAll": "See All Providers",
+    "aiProviderSetting": "Provider Settings for {providerName}",
+    "aiProviderSettingDescription": "Configure this AI provider",
+    "aiProviderGeneral": "General",
+    "aiProviderGeneralDescription": "Basic settings for this provider",
+    "aiProviderConfiguration": "Configuration",
+    "aiProviderConfigurationDescription": "Network routing and authentication for this provider",
+    "aiProviderNetworkSettings": "Network Settings",
+    "aiProviderNetworkSettingsDescription": "Choose how traffic reaches this provider",
+    "aiProviderAuthSettings": "Authentication",
+    "aiProviderAuthSettingsDescription": "Configure how this provider authenticates requests to its upstream URL",
+    "aiProviderType": "Provider Type",
+    "aiProviderTypeSearch": "Search providers...",
+    "aiProviderTypeNotFound": "No provider type found",
+    "aiProviderTypeOpenai": "OpenAI",
+    "aiProviderTypeAnthropic": "Anthropic",
+    "aiProviderTypeGoogleGemini": "Google Gemini",
+    "aiProviderTypeVertexAi": "Vertex AI",
+    "aiProviderTypeBedrock": "Amazon Bedrock",
+    "aiProviderTypeMicrosoftFoundry": "Microsoft Foundry",
+    "aiProviderTypeOpenRouter": "OpenRouter",
+    "aiProviderTypeVercelAiGateway": "Vercel AI Gateway",
+    "aiProviderTypeCustom": "Custom",
+    "aiProviderTypeOpenaiDescription": "OpenAI API with default upstream URL",
+    "aiProviderTypeAnthropicDescription": "Anthropic API with default upstream URL",
+    "aiProviderTypeGoogleGeminiDescription": "Google Gemini generateContent API",
+    "aiProviderTypeVertexAiDescription": "Google Vertex AI; upstream URL required",
+    "aiProviderTypeBedrockDescription": "Amazon Bedrock Runtime",
+    "aiProviderTypeMicrosoftFoundryDescription": "Microsoft Foundry; upstream URL required",
+    "aiProviderTypeOpenRouterDescription": "OpenRouter API",
+    "aiProviderTypeVercelAiGatewayDescription": "Vercel AI Gateway",
+    "aiProviderTypeCustomDescription": "Bring your own OpenAI-compatible endpoint or route via Pangolin targets",
+    "aiProviderUpstreamUrl": "Upstream URL",
+    "aiProviderUpstreamUrlDescription": "Base URL for the provider API",
+    "aiProviderUpstreamUrlOptionalDescription": "Leave blank to use the default upstream URL for this provider",
+    "aiProviderEffectiveUpstreamUrl": "Effective Upstream URL",
+    "aiProviderApiKey": "API Key",
+    "aiProviderApiKeyDescription": "API key used to authenticate requests to this provider",
+    "aiProviderCustomHeadersDescription": "Headers sent on every request to this provider. Newline separated: Header-Name: value",
+    "aiProviderApiKeyLastChars": "API Key",
+    "aiProviderAuthType": "Auth Type",
+    "aiProviderAuthTypeSearch": "Search auth types...",
+    "aiProviderAuthTypeNotFound": "No auth type found",
+    "aiProviderAuthTypeBearer": "Bearer",
+    "aiProviderAuthTypeBearerDescription": "Authorization: Bearer key. Used by OpenAI and most providers",
+    "aiProviderAuthTypeXApiKey": "x-api-key",
+    "aiProviderAuthTypeXApiKeyDescription": "x-api-key header. Used by Anthropic",
+    "aiProviderAuthTypeXGoogApiKey": "x-goog-api-key",
+    "aiProviderAuthTypeXGoogApiKeyDescription": "x-goog-api-key header. Used by Google Gemini",
+    "aiProviderAuthTypeHec": "Splunk HEC",
+    "aiProviderAuthTypeHecDescription": "Authorization: Splunk key. Used by Splunk HTTP Event Collector",
+    "aiProviderAuthTypeCfAigAuthorization": "Cloudflare AI Gateway",
+    "aiProviderAuthTypeCfAigAuthorizationDescription": "cf-aig-authorization: Bearer key. Used by Cloudflare AI Gateway",
+    "aiProviderAuthTypeNone": "No Auth",
+    "aiProviderAuthTypePassthrough": "Passthrough",
+    "aiProviderAuthTypeDescription": "How the upstream API authenticates requests",
+    "aiProviderAuthTypePassthroughDescription": "Forward the caller's API key headers to the upstream",
+    "aiProviderAuthTypeNoneDescription": "Do not send authentication headers to the upstream",
+    "aiProviderRoutingMode": "Routing Mode",
+    "aiProviderRoutingModeDescription": "Send traffic to an upstream URL or to HTTP targets on your sites",
+    "aiProviderRoutingModeUrl": "Upstream URL",
+    "aiProviderRoutingModeUrlDescription": "Call a public or private API base URL",
+    "aiProviderRoutingModeTarget": "Site Targets",
+    "aiProviderRoutingModeTargetDescription": "Route through targets on your sites",
+    "aiProviderRoutingModeTargetNote": "After creating this provider, configure site targets on the Network Settings tab.",
+    "aiProviderTargetNoOne": "This provider doesn't have any targets. Add a target to route requests through your sites.",
+    "aiProviderSkipTlsVerification": "Skip TLS Verification",
+    "aiProviderSkipTlsVerificationDescription": "Disable TLS certificate verification for the upstream connection",
+    "aiProviderBudget": "Budget",
+    "aiProviderBudgetDescription": "Optional spending or token budget for this provider",
+    "aiProviderBudgetAmount": "Budget Amount",
+    "aiProviderBudgetUnit": "Budget Unit",
+    "aiProviderBudgetUnitUsd": "USD",
+    "aiProviderBudgetUnitTokens": "Tokens",
+    "aiProviderEnabled": "Enabled",
+    "aiProviderEnabledDescription": "Disable to stop using this provider without deleting it",
+    "aiProviderErrorCreate": "Failed to create AI provider",
+    "aiProviderErrorUpdate": "Failed to update AI provider",
+    "aiProviderErrorDelete": "Failed to delete AI provider",
+    "aiProviderErrorLoad": "Failed to load AI provider",
+    "aiProviderErrorUpstreamUrlInvalid": "Enter a valid upstream URL",
+    "aiProviderErrorUpstreamUrlRequired": "Upstream URL is required for this provider",
+    "aiProviderErrorAuthTypeRequired": "Auth type is required",
+    "aiProviderErrorApiKeyRequired": "API key is required",
+    "aiProviderErrorRoutingModeTarget": "Site targets routing is only available for custom providers",
+    "aiProviderErrorCapabilitiesRequired": "Select at least one API capability",
+    "aiProviderCapabilities": "API Capabilities",
+    "aiProviderCapabilitiesDescription": "Select which API formats this provider can handle. Known providers start with recommended defaults.",
+    "aiProviderCapabilitiesCustomDescription": "Select which API formats this custom provider can handle",
+    "aiProviderCapabilitiesSelect": "Select capabilities",
+    "aiProviderCapabilitiesEmpty": "No capabilities found",
+    "aiProviderCapabilitiesSearch": "Search capabilities...",
+    "aiCapabilityOpenaiChat": "OpenAI Chat Completions",
+    "aiCapabilityOpenaiChatDescription": "Supports /v1/chat/completions",
+    "aiCapabilityOpenaiResponses": "OpenAI Responses",
+    "aiCapabilityOpenaiResponsesDescription": "Supports /v1/responses",
+    "aiCapabilityAnthropicMessages": "Anthropic Messages",
+    "aiCapabilityAnthropicMessagesDescription": "Supports /v1/messages",
+    "aiCapabilityGeminiGenerateContent": "Gemini Generate Content",
+    "aiCapabilityGeminiGenerateContentDescription": "Supports the direct Gemini API",
+    "aiCapabilityBedrockModelInvoke": "Bedrock Model Invoke",
+    "aiCapabilityBedrockModelInvokeDescription": "Supports Amazon Bedrock InvokeModel",
+    "aiCapabilityGoogleGenerateContent": "Vertex Generate Content",
+    "aiCapabilityGoogleGenerateContentDescription": "Supports Vertex AI Gemini format",
+    "aiCapabilityGoogleRawPredict": "Vertex Raw Predict",
+    "aiCapabilityGoogleRawPredictDescription": "Supports Vertex AI rawPredict for Anthropic models",
+    "aiCapabilityBedrockConverse": "Bedrock Converse",
+    "aiCapabilityBedrockConverseDescription": "Supports Amazon Bedrock Converse API",
+    "aiProviderCreated": "AI provider created",
+    "aiProviderUpdated": "AI provider updated",
+    "aiProviderDeleted": "AI provider deleted",
+    "aiProviderDelete": "Delete Provider",
+    "aiProviderDeleteConfirm": "Delete Provider",
+    "aiProviderQuestionRemove": "Are you sure you want to delete this AI provider?",
+    "aiProviderMessageRemove": "This will permanently delete the provider and its models and targets. This cannot be undone.",
+    "aiProviderErrorNoUpdate": "AI provider is not available to update",
+    "aiProviderModels": "Models",
+    "aiProviderModelsDescription": "Define allow and block patterns for this provider. Requests must match an allow pattern and must not match a block pattern. Use * and ? as wildcards (for example gpt-4* or claude-?). An empty allow list denies all models.",
+    "aiProviderModelsPlaceholder": "Model name or pattern (e.g. gpt-4*)",
+    "aiProviderModelsAllow": "Allow List",
+    "aiProviderModelsAllowDescription": "Models that may be used through this provider. Empty means deny all.",
+    "aiProviderModelsAllowPlaceholder": "Allowed model or pattern (e.g. gpt-4*)",
+    "aiProviderModelsBlock": "Block List",
+    "aiProviderModelsBlockDescription": "Models to deny even if they match an allow pattern.",
+    "aiProviderModelsBlockPlaceholder": "Blocked model or pattern (e.g. gpt-4o-mini)",
+    "aiProviderModelsOverlapError": "These patterns cannot be on both lists: {keys}",
+    "aiProviderModelsUpdated": "Models updated",
+    "aiProviderModelsErrorUpdate": "Failed to update models",
+    "aiResourceProviders": "Providers",
+    "aiResourceProvidersDescription": "Choose which AI providers this inference resource can use",
+    "aiResourceProvidersHelp": "Attach providers and choose inherit (use each provider's lists) or select (pick an allow list for this resource). Allow patterns that conflict across attached providers are not allowed.",
+    "aiResourceProvidersSelect": "Select providers",
+    "aiResourceProvidersEmpty": "No AI providers found",
+    "aiResourceProvidersNoneAttached": "No providers attached yet.",
+    "aiResourceProvidersAdd": "Add provider",
+    "aiResourceProvidersRemove": "Remove provider",
+    "aiResourceProviderToggleEnabled": "Enable or disable this provider on the resource",
+    "aiResourceProviderDisabled": "Disabled",
+    "aiResourceProvidersUpdated": "Providers updated",
+    "aiResourceProvidersErrorUpdate": "Failed to update providers",
+    "aiResourceProviderEditDescription": "Choose how this provider's models are exposed on this resource.",
+    "aiResourceProviderMode": "Access mode",
+    "aiResourceProviderModeInherit": "Inherit",
+    "aiResourceProviderModeSelect": "Select",
+    "aiResourceProviderModeSelectSummary": "Select · {count} models",
+    "aiResourceProviderModeInheritHelp": "Use this provider's allow and block lists as configured on the provider.",
+    "aiResourceProviderModeSelectHelp": "Choose a subset of this provider's allow-list models for this resource.",
+    "aiResourceProviderAllowModels": "Allow list",
+    "aiResourceProviderAllowModelsSelect": "Select models",
+    "aiResourceProviderAllowModelsSearch": "Search models...",
+    "aiResourceProviderAllowModelsEmpty": "No models found",
+    "aiResourceProviderAllowModelsHelp": "Only models from this provider's allow list can be selected.",
+    "aiResourceAliasRequired": "Alias is required for inference resources",
+    "aiResourceDomainConfiguration": "Domain configuration",
+    "aiResourceDomainConfigurationDescription": "Choose the domain clients will use to reach this inference resource.",
    "sidebarApiKeys": "API Keys",
    "sidebarProvisioning": "Provisioning",
    "sidebarSettings": "Settings",
@@ -2246,6 +2421,7 @@
    "requireDeviceApproval": "Require Device Approvals",
    "requireDeviceApprovalDescription": "Users with this role need new devices approved by an admin before they can connect and access resources.",
    "sshSettings": "SSH Settings",
+    "inferenceSettings": "Inference Settings",
    "sshAccess": "SSH Access",
    "rdpSettings": "RDP Settings",
    "vncSettings": "VNC Settings",
@@ -2401,6 +2577,7 @@
    "editInternalResourceDialogModeCidr": "CIDR",
    "editInternalResourceDialogModeHttp": "HTTP",
    "editInternalResourceDialogModeHttps": "HTTPS",
+    "editInternalResourceDialogModeInference": "Inference",
    "editInternalResourceDialogModeSsh": "SSH",
    "editInternalResourceDialogScheme": "Scheme",
    "editInternalResourceDialogEnableSsl": "Enable TLS",
@@ -2460,6 +2637,7 @@
    "createInternalResourceDialogModeHttp": "HTTP",
    "createInternalResourceDialogModeHttps": "HTTPS",
    "createInternalResourceDialogModeSsh": "SSH",
+    "createInternalResourceDialogModeInference": "Inference",
    "scheme": "Scheme",
    "createInternalResourceDialogScheme": "Scheme",
    "createInternalResourceDialogEnableSsl": "Enable TLS",
@@ -65,6 +65,7 @@
                "express": "5.2.1",
                "express-rate-limit": "8.5.2",
                "glob": "13.0.6",
+                "gpt-tokenizer": "^3.4.0",
                "helmet": "8.2.0",
                "http-errors": "2.0.1",
                "input-otp": "1.4.2",
@@ -12760,6 +12761,12 @@
                "url": "https://github.com/sponsors/ljharb"
            }
        },
+        "node_modules/gpt-tokenizer": {
+            "version": "3.4.0",
+            "resolved": "https://registry.npmjs.org/gpt-tokenizer/-/gpt-tokenizer-3.4.0.tgz",
+            "integrity": "sha512-wxFLnhIXTDjYebd9A9pGl3e31ZpSypbpIJSOswbgop5jLte/AsZVDvjlbEuVFlsqZixVKqbcoNmRlFDf6pz/UQ==",
+            "license": "MIT"
+        },
        "node_modules/graceful-fs": {
            "version": "4.2.11",
            "resolved": "https://registry.npmjs.org/graceful-fs/-/graceful-fs-4.2.11.tgz",
@@ -33,9 +33,9 @@
    },
    "dependencies": {
        "@asteasolutions/zod-to-openapi": "8.5.0",
+        "@aws-sdk/client-s3": "3.1056.0",
        "@devolutions/iron-remote-desktop": "https://static.pangolin.net/packages/devolutions-iron-remote-desktop-0.0.0.tgz",
        "@devolutions/iron-remote-desktop-rdp": "https://static.pangolin.net/packages/devolutions-iron-remote-desktop-rdp-0.0.1.tgz",
-        "@aws-sdk/client-s3": "3.1056.0",
        "@headlessui/react": "2.2.10",
        "@hookform/resolvers": "5.4.0",
        "@monaco-editor/react": "4.7.0",
@@ -88,6 +88,7 @@
        "express": "5.2.1",
        "express-rate-limit": "8.5.2",
        "glob": "13.0.6",
+        "gpt-tokenizer": "^3.4.0",
        "helmet": "8.2.0",
        "http-errors": "2.0.1",
        "input-otp": "1.4.2",
@@ -0,0 +1,39 @@
+import express from "express";
+import helmet from "helmet";
+import cors from "cors";
+import config from "@server/lib/config";
+import logger from "@server/logger";
+import {
+    errorHandlerMiddleware,
+    notFoundMiddleware
+} from "@server/middlewares";
+import { createAiGatewayRouter } from "@server/routers/aiGateway";
+
+const aiGatewayPort = config.getRawConfig().server.ai_gateway_port;
+
+export function createAiGatewayServer() {
+    const aiGatewayServer = express();
+
+    const trustProxy = config.getRawConfig().server.trust_proxy;
+    if (trustProxy) {
+        aiGatewayServer.set("trust proxy", trustProxy);
+    }
+
+    aiGatewayServer.use(helmet());
+    aiGatewayServer.use(cors());
+    aiGatewayServer.use(express.json());
+
+    aiGatewayServer.use(createAiGatewayRouter());
+
+    aiGatewayServer.use(notFoundMiddleware);
+    aiGatewayServer.use(errorHandlerMiddleware);
+
+    aiGatewayServer.listen(aiGatewayPort, (err?: any) => {
+        if (err) throw err;
+        logger.info(
+            `AI gateway server is running on http://localhost:${aiGatewayPort}`
+        );
+    });
+
+    return aiGatewayServer;
+}
@@ -50,6 +50,8 @@ export enum ActionsEnum {
    setResourceUsers = "setResourceUsers",
    setResourceRoles = "setResourceRoles",
    listResourceUsers = "listResourceUsers",
+    listResourceAiModels = "listResourceAiModels",
+    setResourceAiModels = "setResourceAiModels",
    // removeRoleSite = "removeRoleSite",
    // addRoleAction = "addRoleAction",
    // removeRoleAction = "removeRoleAction",
@@ -182,7 +184,17 @@ export enum ActionsEnum {
    setResourcePolicyHeaderAuth = "setResourcePolicyHeaderAuth",
    setResourcePolicyWhitelist = "setResourcePolicyWhitelist",
    setResourcePolicyRules = "setResourcePolicyRules",
-    createOrgWideLauncherView = "createOrgWideLauncherView"
+    createOrgWideLauncherView = "createOrgWideLauncherView",
+    createAiProvider = "createAiProvider",
+    deleteAiProvider = "deleteAiProvider",
+    getAiProvider = "getAiProvider",
+    listAiProviders = "listAiProviders",
+    updateAiProvider = "updateAiProvider",
+    createAiModel = "createAiModel",
+    deleteAiModel = "deleteAiModel",
+    getAiModel = "getAiModel",
+    listAiModels = "listAiModels",
+    updateAiModel = "updateAiModel"
 }

 export async function checkUserActionPermission(
@@ -3,6 +3,7 @@ import { InferSelectModel, sql } from "drizzle-orm";
 import {
    bigint,
    boolean,
+    check,
    index,
    integer,
    pgTable,
@@ -99,6 +100,7 @@ export const sites = pgTable(
        name: varchar("name").notNull(),
        pubKey: varchar("pubKey"),
        subnet: varchar("subnet"),
+        exitNodeSubnet: text("exitNodeSubnet"), // this is the subnet when connecting to an exit node and INCLUDES THE CIDR
        megabytesIn: real("bytesIn").default(0),
        megabytesOut: real("bytesOut").default(0),
        lastBandwidthUpdate: varchar("lastBandwidthUpdate"),
@@ -194,7 +196,12 @@ export const resources = pgTable(
        postAuthPath: text("postAuthPath"),
        health: varchar("health").default("unknown"), // "healthy", "unhealthy", "unknown"
        wildcard: boolean("wildcard").notNull().default(false),
-        mode: text("mode").default("http").notNull(), // rdp, ssh, http, vnc
+        mode: text("mode")
+            .default("http")
+            .$type<
+                "rdp" | "ssh" | "http" | "vnc" | "inference" | "tcp" | "udp"
+            >()
+            .notNull(),
        pamMode: varchar("pamMode", { length: 32 })
            .$type<"passthrough" | "push">()
            .default("passthrough"),
@@ -215,6 +222,41 @@ export const resources = pgTable(
    ]
 );

+export const resourceAiProviders = pgTable(
+    "resourceAiProviders",
+    {
+        resourceId: integer("resourceId")
+            .notNull()
+            .references(() => resources.resourceId, { onDelete: "cascade" }),
+        providerId: integer("providerId")
+            .notNull()
+            .references(() => aiProviders.providerId, { onDelete: "cascade" }),
+        accessMode: varchar("accessMode")
+            .$type<"inherit" | "select">()
+            .notNull()
+            .default("inherit"),
+        enabled: boolean("enabled").notNull().default(true)
+    },
+    (t) => [primaryKey({ columns: [t.resourceId, t.providerId] })]
+);
+
+export const resourceAiModels = pgTable(
+    "resourceAiModels",
+    {
+        resourceId: integer("resourceId")
+            .notNull()
+            .references(() => resources.resourceId, { onDelete: "cascade" }),
+        modelId: integer("modelId")
+            .notNull()
+            .references(() => aiModels.modelId, { onDelete: "cascade" }),
+        listType: varchar("listType")
+            .$type<"allow" | "block">()
+            .notNull()
+            .default("allow")
+    },
+    (t) => [primaryKey({ columns: [t.resourceId, t.modelId] })]
+);
+
 export const labels = pgTable("labels", {
    labelId: serial("labelId").primaryKey(),
    name: varchar("name").notNull(),
@@ -317,11 +359,18 @@ export const targets = pgTable(
    "targets",
    {
        targetId: serial("targetId").primaryKey(),
-        resourceId: integer("resourceId")
-            .references(() => resources.resourceId, {
+        resourceId: integer("resourceId").references(
+            () => resources.resourceId,
+            {
                onDelete: "cascade"
-            })
-            .notNull(),
+            }
+        ),
+        providerId: integer("providerId").references(
+            () => aiProviders.providerId,
+            {
+                onDelete: "cascade"
+            }
+        ),
        siteId: integer("siteId")
            .references(() => sites.siteId, {
                onDelete: "cascade"
@@ -345,6 +394,7 @@ export const targets = pgTable(
    },
    (t) => [
        index("idx_targets_resourceid_siteid").on(t.resourceId, t.siteId),
+        index("idx_targets_providerid_siteid").on(t.providerId, t.siteId),
        index("idx_targets_site_enabled_priority_target_resource")
            .on(t.siteId, t.priority.desc(), t.targetId, t.resourceId)
            .where(sql`${t.enabled} = true`)
@@ -424,11 +474,14 @@ export const siteResources = pgTable(
                onDelete: "restrict"
            }
        ),
+        requiresExitNodeConnection: boolean("requiresExitNodeConnection")
+            .notNull()
+            .default(false),
        niceId: varchar("niceId").notNull(),
        name: varchar("name").notNull(),
        ssl: boolean("ssl").notNull().default(false),
        mode: varchar("mode")
-            .$type<"host" | "cidr" | "http" | "ssh">()
+            .$type<"host" | "cidr" | "http" | "ssh" | "inference">()
            .notNull(), // "host" | "cidr" | "http"
        scheme: varchar("scheme").$type<"http" | "https">(), // only for when we are doing https or http mode
        proxyPort: integer("proxyPort"), // only for port mode
@@ -463,6 +516,45 @@ export const siteResources = pgTable(
    (t) => [index("idx_siteresources_orgid_niceid").on(t.orgId, t.niceId)]
 );

+export const siteResourceAiProviders = pgTable(
+    "siteResourceAiProviders",
+    {
+        siteResourceId: integer("siteResourceId")
+            .notNull()
+            .references(() => siteResources.siteResourceId, {
+                onDelete: "cascade"
+            }),
+        providerId: integer("providerId")
+            .notNull()
+            .references(() => aiProviders.providerId, { onDelete: "cascade" }),
+        accessMode: varchar("accessMode")
+            .$type<"inherit" | "select">()
+            .notNull()
+            .default("inherit"),
+        enabled: boolean("enabled").notNull().default(true)
+    },
+    (t) => [primaryKey({ columns: [t.siteResourceId, t.providerId] })]
+);
+
+export const siteResourceAiModels = pgTable(
+    "siteResourceAiModels",
+    {
+        siteResourceId: integer("siteResourceId")
+            .notNull()
+            .references(() => siteResources.siteResourceId, {
+                onDelete: "cascade"
+            }),
+        modelId: integer("modelId")
+            .notNull()
+            .references(() => aiModels.modelId, { onDelete: "cascade" }),
+        listType: varchar("listType")
+            .$type<"allow" | "block">()
+            .notNull()
+            .default("allow")
+    },
+    (t) => [primaryKey({ columns: [t.siteResourceId, t.modelId] })]
+);
+
 export const networks = pgTable(
    "networks",
    {
@@ -1181,7 +1273,7 @@ export const clients = pgTable(
        olmId: text("olmId"), // to lock it to a specific olm optionally
        name: varchar("name").notNull(),
        pubKey: varchar("pubKey"),
-        subnet: varchar("subnet").notNull(),
+        exitNodeSubnet: varchar("exitNodeSubnet").notNull(), // INCLUDES THE CIDR
        megabytesIn: real("bytesIn"),
        megabytesOut: real("bytesOut"),
        lastBandwidthUpdate: varchar("lastBandwidthUpdate"),
@@ -1540,6 +1632,140 @@ export const statusHistory = pgTable(
    ]
 );

+export const aiProviders = pgTable("aiProviders", {
+    providerId: serial("providerId").primaryKey(),
+    orgId: varchar("orgId")
+        .notNull()
+        .references(() => orgs.orgId, { onDelete: "cascade" }),
+    name: varchar("name").notNull(),
+    type: varchar("type")
+        .$type<
+            | "openai"
+            | "anthropic"
+            | "googleGemini"
+            | "vertexAi"
+            | "bedrock"
+            | "microsoftFoundry"
+            | "openRouter"
+            | "vercelAiGateway"
+            | "custom"
+        >()
+        .notNull(),
+    upstreamUrl: text("upstreamUrl"),
+    apiKey: text("apiKey"),
+    apiKeyLastChars: varchar("apiKeyLastChars"),
+    authType: varchar("authType")
+        .$type<
+            | "bearer"
+            | "x-api-key"
+            | "x-goog-api-key"
+            | "hec"
+            | "cf-aig-authorization"
+            | "none"
+            | "passthrough"
+        >()
+        .notNull(),
+    routingMode: varchar("routingMode")
+        .$type<"url" | "target">()
+        .notNull()
+        .default("url"),
+    capabilities: text("capabilities").notNull().default("[]"),
+    headers: text("headers"), // JSON array of { name, value }
+    skipTlsVerification: boolean("skipTlsVerification")
+        .notNull()
+        .default(false),
+    enabled: boolean("enabled").notNull().default(true),
+    createdAt: bigint("createdAt", { mode: "number" }).notNull(),
+    updatedAt: bigint("updatedAt", { mode: "number" }).notNull()
+});
+
+export const aiModels = pgTable(
+    "aiModels",
+    {
+        modelId: serial("modelId").primaryKey(),
+        providerId: integer("providerId")
+            .notNull()
+            .references(() => aiProviders.providerId, { onDelete: "cascade" }),
+        modelKey: varchar("modelKey").notNull(),
+        name: varchar("name").notNull(),
+        listType: varchar("listType")
+            .$type<"allow" | "block">()
+            .notNull()
+            .default("allow"),
+        enabled: boolean("enabled").notNull().default(true),
+        createdAt: bigint("createdAt", { mode: "number" }).notNull(),
+        updatedAt: bigint("updatedAt", { mode: "number" }).notNull()
+    },
+    (t) => [unique("ai_model_provider_key_uniq").on(t.providerId, t.modelKey)]
+);
+
+export const aiBudgets = pgTable(
+    "aiBudgets",
+    {
+        budgetId: serial("budgetId").primaryKey(),
+        orgId: varchar("orgId")
+            .notNull()
+            .references(() => orgs.orgId, { onDelete: "cascade" }),
+        providerId: integer("providerId").references(
+            () => aiProviders.providerId,
+            { onDelete: "cascade" }
+        ),
+        modelId: integer("modelId").references(() => aiModels.modelId, {
+            onDelete: "cascade"
+        }),
+        resourceId: integer("resourceId").references(
+            () => resources.resourceId,
+            { onDelete: "cascade" }
+        ),
+        siteResourceId: integer("siteResourceId").references(
+            () => siteResources.siteResourceId,
+            { onDelete: "cascade" }
+        ),
+        amount: real("amount").notNull(),
+        unit: varchar("unit").$type<"usd" | "tokens">().notNull(),
+        period: varchar("period")
+            .$type<"monthly">()
+            .notNull()
+            .default("monthly"),
+        enforcement: varchar("enforcement")
+            .$type<"hard" | "soft">()
+            .notNull()
+            .default("hard"),
+        enabled: boolean("enabled").notNull().default(true),
+        createdAt: bigint("createdAt", { mode: "number" }).notNull(),
+        updatedAt: bigint("updatedAt", { mode: "number" }).notNull()
+    },
+    (t) => [
+        check(
+            "ai_budget_one_scope",
+            sql`(
+                (CASE WHEN ${t.providerId} IS NOT NULL THEN 1 ELSE 0 END) +
+                (CASE WHEN ${t.modelId} IS NOT NULL THEN 1 ELSE 0 END) +
+                (CASE WHEN ${t.resourceId} IS NOT NULL THEN 1 ELSE 0 END) +
+                (CASE WHEN ${t.siteResourceId} IS NOT NULL THEN 1 ELSE 0 END)
+            ) = 1`
+        ),
+        unique("ai_budget_provider_uniq").on(t.providerId),
+        unique("ai_budget_model_uniq").on(t.modelId),
+        unique("ai_budget_resource_uniq").on(t.resourceId),
+        unique("ai_budget_site_resource_uniq").on(t.siteResourceId)
+    ]
+);
+
+export const aiBudgetPeriods = pgTable(
+    "aiBudgetPeriods",
+    {
+        periodId: serial("periodId").primaryKey(),
+        budgetId: integer("budgetId")
+            .notNull()
+            .references(() => aiBudgets.budgetId, { onDelete: "cascade" }),
+        periodStart: bigint("periodStart", { mode: "number" }).notNull(),
+        periodEnd: bigint("periodEnd", { mode: "number" }).notNull(),
+        usedAmount: real("usedAmount").notNull().default(0)
+    },
+    (t) => [unique("ai_budget_period_start_uniq").on(t.budgetId, t.periodStart)]
+);
+
 export type Org = InferSelectModel<typeof orgs>;
 export type User = InferSelectModel<typeof users>;
 export type Site = InferSelectModel<typeof sites>;
@@ -1624,3 +1850,13 @@ export type ResourcePolicy = InferSelectModel<typeof resourcePolicies>;
 export type RolePolicy = InferSelectModel<typeof rolePolicies>;
 export type UserPolicy = InferSelectModel<typeof userPolicies>;
 export type ResourcePolicyRule = InferSelectModel<typeof resourcePolicyRules>;
+export type AiProvider = InferSelectModel<typeof aiProviders>;
+export type AiModel = InferSelectModel<typeof aiModels>;
+export type AiBudget = InferSelectModel<typeof aiBudgets>;
+export type AiBudgetPeriod = InferSelectModel<typeof aiBudgetPeriods>;
+export type ResourceAiProvider = InferSelectModel<typeof resourceAiProviders>;
+export type SiteResourceAiProvider = InferSelectModel<
+    typeof siteResourceAiProviders
+>;
+export type ResourceAiModel = InferSelectModel<typeof resourceAiModels>;
+export type SiteResourceAiModel = InferSelectModel<typeof siteResourceAiModels>;
@@ -6,13 +6,48 @@ import fs from "fs";
 import { APP_PATH } from "@server/lib/consts";
 import { existsSync, mkdirSync } from "fs";

+// Temporary diagnostic trace for the random better-sqlite3 native aborts
+// (Statement::~Statement -> RemoveEnvironmentCleanupHook assertion). That
+// abort is a hard SIGABRT from native code, so it bypasses uncaughtException/
+// unhandledRejection and can outrun winston's async file transport. This
+// writes every statement text synchronously (fsync'd via appendFileSync) so
+// the statements immediately preceding a crash survive it. better-sqlite3's
+// `verbose` hook fires for BEGIN/SAVEPOINT/RELEASE/COMMIT/ROLLBACK too, since
+// those are just prepared statements under the hood - so this also lets us
+// see if two "transactions" ever overlap on the shared connection.
+// Enable with SQL_TRACE=true; remove once the crash is root-caused.
+function sqlTraceVerbose():
+    | ((message: unknown, ...args: unknown[]) => void)
+    | undefined {
+    if (process.env.SQL_TRACE !== "true") {
+        return undefined;
+    }
+    const traceLogDir = path.join(APP_PATH, "logs");
+    if (!existsSync(traceLogDir)) {
+        mkdirSync(traceLogDir, { recursive: true });
+    }
+    const traceLogPath = path.join(traceLogDir, "sql-trace.log");
+    let seq = 0;
+    return (message: unknown) => {
+        seq += 1;
+        const line = `${new Date().toISOString()} pid=${process.pid} #${seq} ${String(message).replace(/\s+/g, " ").trim()}\n`;
+        try {
+            fs.appendFileSync(traceLogPath, line);
+        } catch {
+            // best-effort diagnostic logging only
+        }
+    };
+}
+
 export const location = path.join(APP_PATH, "db", "db.sqlite");
 export const exists = checkFileExists(location);

 bootstrapVolume();

 function createDb() {
-    const sqlite = new Database(location);
+    const verbose =
+        process.env.QUERY_LOGGING == "true" ? sqlTraceVerbose() : undefined;
+    const sqlite = new Database(location, { verbose });

    if (process.env.ENABLE_SQLITE_WAL_MODE == "true") {
        // Enable WAL mode — allows concurrent readers + single writer, preventing
@@ -1,9 +1,11 @@
 import { randomUUID } from "crypto";
-import { InferSelectModel } from "drizzle-orm";
+import { InferSelectModel, sql } from "drizzle-orm";
 import {
+    check,
    index,
    integer,
    primaryKey,
+    real,
    sqliteTable,
    text,
    unique
@@ -107,7 +109,7 @@ export const sites = sqliteTable("sites", {
    }),
    name: text("name").notNull(),
    pubKey: text("pubKey"),
-    subnet: text("subnet"),
+    exitNodeSubnet: text("exitNodeSubnet"),
    megabytesIn: integer("bytesIn").default(0),
    megabytesOut: integer("bytesOut").default(0),
    lastBandwidthUpdate: text("lastBandwidthUpdate"),
@@ -203,7 +205,10 @@ export const resources = sqliteTable("resources", {
    postAuthPath: text("postAuthPath"),
    health: text("health").default("unknown"), // "healthy", "unhealthy", "unknown"
    wildcard: integer("wildcard", { mode: "boolean" }).notNull().default(false),
-    mode: text("mode").default("http").notNull(), // rdp, ssh, http, vnc
+    mode: text("mode")
+        .default("http")
+        .$type<"rdp" | "ssh" | "http" | "vnc" | "inference" | "tcp" | "udp">()
+        .notNull(), // rdp, ssh, http, vnc, inference
    pamMode: text("pamMode")
        .$type<"passthrough" | "push">()
        .default("passthrough"),
@@ -214,6 +219,41 @@ export const resources = sqliteTable("resources", {
    status: text("status").$type<"pending" | "approved">().default("approved")
 });

+export const resourceAiProviders = sqliteTable(
+    "resourceAiProviders",
+    {
+        resourceId: integer("resourceId")
+            .notNull()
+            .references(() => resources.resourceId, { onDelete: "cascade" }),
+        providerId: integer("providerId")
+            .notNull()
+            .references(() => aiProviders.providerId, { onDelete: "cascade" }),
+        accessMode: text("accessMode")
+            .$type<"inherit" | "select">()
+            .notNull()
+            .default("inherit"),
+        enabled: integer("enabled", { mode: "boolean" }).notNull().default(true)
+    },
+    (t) => [primaryKey({ columns: [t.resourceId, t.providerId] })]
+);
+
+export const resourceAiModels = sqliteTable(
+    "resourceAiModels",
+    {
+        resourceId: integer("resourceId")
+            .notNull()
+            .references(() => resources.resourceId, { onDelete: "cascade" }),
+        modelId: integer("modelId")
+            .notNull()
+            .references(() => aiModels.modelId, { onDelete: "cascade" }),
+        listType: text("listType")
+            .$type<"allow" | "block">()
+            .notNull()
+            .default("allow")
+    },
+    (t) => [primaryKey({ columns: [t.resourceId, t.modelId] })]
+);
+
 export const labels = sqliteTable("labels", {
    labelId: integer("labelId").primaryKey({ autoIncrement: true }),
    name: text("name").notNull(),
@@ -322,11 +362,12 @@ export const clientLabels = sqliteTable(

 export const targets = sqliteTable("targets", {
    targetId: integer("targetId").primaryKey({ autoIncrement: true }),
-    resourceId: integer("resourceId")
-        .references(() => resources.resourceId, {
-            onDelete: "cascade"
-        })
-        .notNull(),
+    resourceId: integer("resourceId").references(() => resources.resourceId, {
+        onDelete: "cascade"
+    }),
+    providerId: integer("providerId").references(() => aiProviders.providerId, {
+        onDelete: "cascade"
+    }),
    siteId: integer("siteId")
        .references(() => sites.siteId, {
            onDelete: "cascade"
@@ -422,10 +463,17 @@ export const siteResources = sqliteTable("siteResources", {
        () => networks.networkId,
        { onDelete: "restrict" }
    ),
+    requiresExitNodeConnection: integer("requiresExitNodeConnection", {
+        mode: "boolean"
+    })
+        .notNull()
+        .default(false),
    niceId: text("niceId").notNull(),
    name: text("name").notNull(),
    ssl: integer("ssl", { mode: "boolean" }).notNull().default(false),
-    mode: text("mode").$type<"host" | "cidr" | "http" | "ssh">().notNull(), // "host" | "cidr" | "http"
+    mode: text("mode")
+        .$type<"host" | "cidr" | "http" | "ssh" | "inference">()
+        .notNull(), // "host" | "cidr" | "http"
    scheme: text("scheme").$type<"http" | "https">(), // only for when we are doing https or http mode
    proxyPort: integer("proxyPort"), // only for port mode
    destinationPort: integer("destinationPort"), // only for port mode
@@ -453,6 +501,45 @@ export const siteResources = sqliteTable("siteResources", {
    status: text("status").$type<"pending" | "approved">().default("approved")
 });

+export const siteResourceAiProviders = sqliteTable(
+    "siteResourceAiProviders",
+    {
+        siteResourceId: integer("siteResourceId")
+            .notNull()
+            .references(() => siteResources.siteResourceId, {
+                onDelete: "cascade"
+            }),
+        providerId: integer("providerId")
+            .notNull()
+            .references(() => aiProviders.providerId, { onDelete: "cascade" }),
+        accessMode: text("accessMode")
+            .$type<"inherit" | "select">()
+            .notNull()
+            .default("inherit"),
+        enabled: integer("enabled", { mode: "boolean" }).notNull().default(true)
+    },
+    (t) => [primaryKey({ columns: [t.siteResourceId, t.providerId] })]
+);
+
+export const siteResourceAiModels = sqliteTable(
+    "siteResourceAiModels",
+    {
+        siteResourceId: integer("siteResourceId")
+            .notNull()
+            .references(() => siteResources.siteResourceId, {
+                onDelete: "cascade"
+            }),
+        modelId: integer("modelId")
+            .notNull()
+            .references(() => aiModels.modelId, { onDelete: "cascade" }),
+        listType: text("listType")
+            .$type<"allow" | "block">()
+            .notNull()
+            .default("allow")
+    },
+    (t) => [primaryKey({ columns: [t.siteResourceId, t.modelId] })]
+);
+
 export const networks = sqliteTable("networks", {
    networkId: integer("networkId").primaryKey({ autoIncrement: true }),
    niceId: text("niceId"),
@@ -599,6 +686,7 @@ export const clients = sqliteTable("clients", {
    pubKey: text("pubKey"),
    olmId: text("olmId"), // to lock it to a specific olm optionally
    subnet: text("subnet").notNull(),
+    exitNodeSubnet: text("exitNodeSubnet"), // this is the subnet when connecting to an exit node
    megabytesIn: integer("bytesIn"),
    megabytesOut: integer("bytesOut"),
    lastBandwidthUpdate: text("lastBandwidthUpdate"),
@@ -1526,6 +1614,132 @@ export const statusHistory = sqliteTable(
    ]
 );

+export const aiProviders = sqliteTable("aiProviders", {
+    providerId: integer("providerId").primaryKey({ autoIncrement: true }),
+    orgId: text("orgId")
+        .notNull()
+        .references(() => orgs.orgId, { onDelete: "cascade" }),
+    name: text("name").notNull(),
+    type: text("type")
+        .$type<
+            | "openai"
+            | "anthropic"
+            | "googleGemini"
+            | "vertexAi"
+            | "bedrock"
+            | "microsoftFoundry"
+            | "openRouter"
+            | "vercelAiGateway"
+            | "custom"
+        >()
+        .notNull(),
+    upstreamUrl: text("upstreamUrl"),
+    apiKey: text("apiKey"),
+    apiKeyLastChars: text("apiKeyLastChars"),
+    authType: text("authType")
+        .$type<
+            | "bearer"
+            | "x-api-key"
+            | "x-goog-api-key"
+            | "hec"
+            | "cf-aig-authorization"
+            | "none"
+            | "passthrough"
+        >()
+        .notNull(),
+    routingMode: text("routingMode")
+        .$type<"url" | "target">()
+        .notNull()
+        .default("url"),
+    capabilities: text("capabilities").notNull().default("[]"),
+    headers: text("headers"), // JSON array of { name, value }
+    skipTlsVerification: integer("skipTlsVerification", { mode: "boolean" })
+        .notNull()
+        .default(false),
+    enabled: integer("enabled", { mode: "boolean" }).notNull().default(true),
+    createdAt: integer("createdAt").notNull(),
+    updatedAt: integer("updatedAt").notNull()
+});
+
+export const aiModels = sqliteTable(
+    "aiModels",
+    {
+        modelId: integer("modelId").primaryKey({ autoIncrement: true }),
+        providerId: integer("providerId")
+            .notNull()
+            .references(() => aiProviders.providerId, { onDelete: "cascade" }),
+        modelKey: text("modelKey").notNull(),
+        name: text("name").notNull(),
+        listType: text("listType")
+            .$type<"allow" | "block">()
+            .notNull()
+            .default("allow"),
+        enabled: integer("enabled", { mode: "boolean" })
+            .notNull()
+            .default(true),
+        createdAt: integer("createdAt").notNull(),
+        updatedAt: integer("updatedAt").notNull()
+    },
+    (t) => [unique("ai_model_provider_key_uniq").on(t.providerId, t.modelKey)]
+);
+
+export const aiBudgets = sqliteTable(
+    "aiBudgets",
+    {
+        budgetId: integer("budgetId").primaryKey({ autoIncrement: true }),
+        orgId: text("orgId")
+            .notNull()
+            .references(() => orgs.orgId, { onDelete: "cascade" }),
+        providerId: integer("providerId").references(
+            () => aiProviders.providerId,
+            { onDelete: "cascade" }
+        ),
+        modelId: integer("modelId").references(() => aiModels.modelId, {
+            onDelete: "cascade"
+        }),
+        resourceId: integer("resourceId").references(
+            () => resources.resourceId,
+            { onDelete: "cascade" }
+        ),
+        siteResourceId: integer("siteResourceId").references(
+            () => siteResources.siteResourceId,
+            { onDelete: "cascade" }
+        ),
+        amount: real("amount").notNull(),
+        unit: text("unit").$type<"usd" | "tokens">().notNull(),
+        period: text("period").$type<"monthly">().notNull().default("monthly"),
+        enforcement: text("enforcement")
+            .$type<"hard" | "soft">()
+            .notNull()
+            .default("hard"),
+        enabled: integer("enabled", { mode: "boolean" })
+            .notNull()
+            .default(true),
+        createdAt: integer("createdAt").notNull(),
+        updatedAt: integer("updatedAt").notNull()
+    },
+    (t) => [
+        unique("ai_budget_provider_uniq").on(t.providerId),
+        unique("ai_budget_model_uniq").on(t.modelId),
+        unique("ai_budget_resource_uniq").on(t.resourceId),
+        unique("ai_budget_site_resource_uniq").on(t.siteResourceId)
+    ]
+);
+
+export const aiBudgetPeriods = sqliteTable(
+    "aiBudgetPeriods",
+    {
+        periodId: integer("periodId").primaryKey({ autoIncrement: true }),
+        budgetId: integer("budgetId")
+            .notNull()
+            .references(() => aiBudgets.budgetId, { onDelete: "cascade" }),
+        periodStart: integer("periodStart").notNull(),
+        periodEnd: integer("periodEnd").notNull(),
+        usedAmount: real("usedAmount").notNull().default(0)
+    },
+    (t) => [unique("ai_budget_period_start_uniq").on(t.budgetId, t.periodStart)]
+);
+
 export type Org = InferSelectModel<typeof orgs>;
 export type User = InferSelectModel<typeof users>;
 export type Site = InferSelectModel<typeof sites>;
@@ -1608,3 +1822,13 @@ export type ResourcePolicyHeaderAuth = InferSelectModel<
 >;
 export type RolePolicy = InferSelectModel<typeof rolePolicies>;
 export type UserPolicy = InferSelectModel<typeof userPolicies>;
+export type AiProvider = InferSelectModel<typeof aiProviders>;
+export type AiModel = InferSelectModel<typeof aiModels>;
+export type AiBudget = InferSelectModel<typeof aiBudgets>;
+export type AiBudgetPeriod = InferSelectModel<typeof aiBudgetPeriods>;
+export type ResourceAiProvider = InferSelectModel<typeof resourceAiProviders>;
+export type SiteResourceAiProvider = InferSelectModel<
+    typeof siteResourceAiProviders
+>;
+export type ResourceAiModel = InferSelectModel<typeof resourceAiModels>;
+export type SiteResourceAiModel = InferSelectModel<typeof siteResourceAiModels>;
@@ -5,10 +5,13 @@ import { runSetupFunctions } from "./setup";
 import { createApiServer } from "./apiServer";
 import { createNextServer } from "./nextServer";
 import { createInternalServer } from "./internalServer";
+import { createAiGatewayServer } from "./aiGatewayServer";
 import { createIntegrationApiServer } from "./integrationApiServer";
 import {
    ApiKey,
    ApiKeyOrg,
+    AiModel,
+    AiProvider,
    RemoteExitNode,
    Session,
    SiteResource,
@@ -47,6 +50,7 @@ async function startServers() {
    // Start all servers
    const apiServer = createApiServer();
    const internalServer = createInternalServer();
+    const aiGatewayServer = createAiGatewayServer();

    const nextServer = await createNextServer();
    if (config.getRawConfig().traefik.file_mode) {
@@ -65,6 +69,7 @@ async function startServers() {
        apiServer,
        nextServer,
        internalServer,
+        aiGatewayServer,
        integrationServer
    };
 }
@@ -83,6 +88,8 @@ declare global {
            userOrgIds?: string[];
            remoteExitNode?: RemoteExitNode;
            siteResource?: SiteResource;
+            aiProvider?: AiProvider;
+            aiModel?: AiModel;
            orgPolicyAllowed?: boolean;
        }
    }
@@ -0,0 +1,283 @@
+import type { Request } from "express";
+
+export const AI_CAPABILITIES = [
+    "openai_chat",
+    "openai_responses",
+    "anthropic_messages",
+    "gemini_generate_content",
+    "bedrock_model_invoke",
+    "google_generate_content",
+    "google_raw_predict",
+    "bedrock_converse"
+] as const;
+
+export type AiCapability = (typeof AI_CAPABILITIES)[number];
+
+export type AiCapabilityRoute = {
+    method: "POST";
+    path: string;
+};
+
+export type AiCapabilityDefinition = {
+    id: AiCapability;
+    routes: AiCapabilityRoute[];
+    extractModel: (req: Request) => string | undefined;
+    resolveUpstreamUrl: (
+        baseUrl: string,
+        req: Request,
+        model: string
+    ) => string;
+    isStreaming: (req: Request, contentType: string) => boolean;
+};
+
+function bodyModel(req: Request): string | undefined {
+    return typeof req.body?.model === "string" ? req.body.model : undefined;
+}
+
+function paramModel(req: Request): string | undefined {
+    const model = req.params?.model;
+    return typeof model === "string" && model.length > 0 ? model : undefined;
+}
+
+export function joinUpstreamUrl(baseUrl: string, path: string): string {
+    const base = baseUrl.replace(/\/+$/, "");
+    let suffix = path.startsWith("/") ? path : `/${path}`;
+
+    let basePathname = "/";
+    try {
+        basePathname = new URL(base).pathname.replace(/\/+$/, "") || "/";
+    } catch {
+        // Fall through with "/" non-absolute bases are not expected in
+        // production, but keep joining usable for malformed input.
+    }
+
+    if (basePathname !== "/") {
+        const baseSegs = basePathname.split("/").filter(Boolean);
+        const pathSegs = suffix.split("/").filter(Boolean);
+        const max = Math.min(baseSegs.length, pathSegs.length);
+        let overlap = 0;
+        for (let n = max; n >= 1; n--) {
+            const baseSuffix = baseSegs.slice(-n);
+            const pathPrefix = pathSegs.slice(0, n);
+            if (baseSuffix.every((seg, i) => seg === pathPrefix[i])) {
+                overlap = n;
+                break;
+            }
+        }
+        if (overlap > 0) {
+            const remaining = pathSegs.slice(overlap);
+            suffix = remaining.length > 0 ? `/${remaining.join("/")}` : "/";
+        }
+    }
+
+    if (suffix === "/") {
+        return base;
+    }
+
+    return `${base}${suffix}`;
+}
+
+function pathFromRequest(req: Request): string {
+    const raw = req.originalUrl || req.url || req.path;
+    return raw.startsWith("/") ? raw : `/${raw}`;
+}
+
+function bodyRequestsStream(req: Request): boolean {
+    return req.body?.stream === true;
+}
+
+function contentTypeIsSse(contentType: string): boolean {
+    return contentType.includes("text/event-stream");
+}
+
+function contentTypeIsAmazonEventStream(contentType: string): boolean {
+    return contentType.includes("application/vnd.amazon.eventstream");
+}
+
+function pathIncludes(req: Request, fragment: string): boolean {
+    return pathFromRequest(req).includes(fragment);
+}
+
+function isBodyOrSseStreaming(req: Request, contentType: string): boolean {
+    return bodyRequestsStream(req) || contentTypeIsSse(contentType);
+}
+
+function isGeminiStyleStreaming(req: Request, contentType: string): boolean {
+    return (
+        pathIncludes(req, "streamGenerateContent") ||
+        pathIncludes(req, "alt=sse") ||
+        contentTypeIsSse(contentType)
+    );
+}
+
+export const AI_CAPABILITY_DEFS: Record<AiCapability, AiCapabilityDefinition> =
+    {
+        openai_chat: {
+            id: "openai_chat",
+            routes: [
+                { method: "POST", path: "/v1/chat/completions" },
+                { method: "POST", path: "/chat/completions" }
+            ],
+            extractModel: bodyModel,
+            resolveUpstreamUrl: (base, req) =>
+                joinUpstreamUrl(base, pathFromRequest(req)),
+            isStreaming: isBodyOrSseStreaming
+        },
+        openai_responses: {
+            id: "openai_responses",
+            routes: [{ method: "POST", path: "/v1/responses" }],
+            extractModel: bodyModel,
+            resolveUpstreamUrl: (base, req) =>
+                joinUpstreamUrl(base, pathFromRequest(req)),
+            isStreaming: isBodyOrSseStreaming
+        },
+        anthropic_messages: {
+            id: "anthropic_messages",
+            routes: [{ method: "POST", path: "/v1/messages" }],
+            extractModel: bodyModel,
+            resolveUpstreamUrl: (base, req) =>
+                joinUpstreamUrl(base, pathFromRequest(req)),
+            isStreaming: isBodyOrSseStreaming
+        },
+        gemini_generate_content: {
+            id: "gemini_generate_content",
+            routes: [
+                {
+                    method: "POST",
+                    path: "/v1beta/models/:model\\:generateContent"
+                },
+                {
+                    method: "POST",
+                    path: "/v1beta/models/:model\\:streamGenerateContent"
+                }
+            ],
+            extractModel: paramModel,
+            resolveUpstreamUrl: (base, req) =>
+                joinUpstreamUrl(base, pathFromRequest(req)),
+            isStreaming: isGeminiStyleStreaming
+        },
+        google_generate_content: {
+            id: "google_generate_content",
+            routes: [
+                {
+                    method: "POST",
+                    // Vertex publisher model generateContent
+                    path: "/v1/projects/:project/locations/:location/publishers/:publisher/models/:model\\:generateContent"
+                },
+                {
+                    method: "POST",
+                    path: "/v1/projects/:project/locations/:location/publishers/:publisher/models/:model\\:streamGenerateContent"
+                }
+            ],
+            extractModel: paramModel,
+            resolveUpstreamUrl: (base, req) =>
+                joinUpstreamUrl(base, pathFromRequest(req)),
+            isStreaming: isGeminiStyleStreaming
+        },
+        google_raw_predict: {
+            id: "google_raw_predict",
+            routes: [
+                {
+                    method: "POST",
+                    path: "/v1/projects/:project/locations/:location/publishers/:publisher/models/:model\\:rawPredict"
+                },
+                {
+                    method: "POST",
+                    path: "/v1/projects/:project/locations/:location/publishers/:publisher/models/:model\\:streamRawPredict"
+                }
+            ],
+            extractModel: paramModel,
+            resolveUpstreamUrl: (base, req) =>
+                joinUpstreamUrl(base, pathFromRequest(req)),
+            isStreaming: (req, contentType) =>
+                pathIncludes(req, "streamRawPredict") ||
+                pathIncludes(req, "alt=sse") ||
+                contentTypeIsSse(contentType)
+        },
+        bedrock_model_invoke: {
+            id: "bedrock_model_invoke",
+            routes: [
+                { method: "POST", path: "/model/:model/invoke" },
+                {
+                    method: "POST",
+                    path: "/model/:model/invoke-with-response-stream"
+                }
+            ],
+            extractModel: paramModel,
+            resolveUpstreamUrl: (base, req) =>
+                joinUpstreamUrl(base, pathFromRequest(req)),
+            isStreaming: (req, contentType) =>
+                pathIncludes(req, "invoke-with-response-stream") ||
+                contentTypeIsAmazonEventStream(contentType) ||
+                contentTypeIsSse(contentType)
+        },
+        bedrock_converse: {
+            id: "bedrock_converse",
+            routes: [
+                { method: "POST", path: "/model/:model/converse" },
+                { method: "POST", path: "/model/:model/converse-stream" }
+            ],
+            extractModel: paramModel,
+            resolveUpstreamUrl: (base, req) =>
+                joinUpstreamUrl(base, pathFromRequest(req)),
+            isStreaming: (req, contentType) =>
+                pathIncludes(req, "converse-stream") ||
+                contentTypeIsAmazonEventStream(contentType) ||
+                contentTypeIsSse(contentType)
+        }
+    };
+
+export function isAiCapability(value: unknown): value is AiCapability {
+    return (
+        typeof value === "string" &&
+        (AI_CAPABILITIES as readonly string[]).includes(value)
+    );
+}
+
+export function parseCapabilities(raw: unknown): AiCapability[] {
+    if (raw == null) {
+        return [];
+    }
+
+    let parsed: unknown = raw;
+    if (typeof raw === "string") {
+        const trimmed = raw.trim();
+        if (!trimmed) {
+            return [];
+        }
+        try {
+            parsed = JSON.parse(trimmed);
+        } catch {
+            return [];
+        }
+    }
+
+    if (!Array.isArray(parsed)) {
+        return [];
+    }
+
+    const out: AiCapability[] = [];
+    const seen = new Set<AiCapability>();
+    for (const item of parsed) {
+        if (isAiCapability(item) && !seen.has(item)) {
+            seen.add(item);
+            out.push(item);
+        }
+    }
+    return out;
+}
+
+export function serializeCapabilities(capabilities: AiCapability[]): string {
+    return JSON.stringify(capabilities);
+}
+
+export function providerHasCapability(
+    capabilities: AiCapability[] | string | null | undefined,
+    capability: AiCapability
+): boolean {
+    const list =
+        typeof capabilities === "string" || capabilities == null
+            ? parseCapabilities(capabilities)
+            : capabilities;
+    return list.includes(capability);
+}
@@ -0,0 +1,82 @@
+import http from "node:http";
+import https from "node:https";
+import { Readable } from "node:stream";
+
+type UpstreamFetchInit = {
+    method: string;
+    headers: Record<string, string>;
+    body?: string;
+    skipTlsVerification?: boolean;
+    signal?: AbortSignal;
+};
+
+const insecureHttpsAgent = new https.Agent({
+    rejectUnauthorized: false,
+    keepAlive: true
+});
+
+export function aiGatewayUpstreamFetch(
+    url: string,
+    init: UpstreamFetchInit
+): Promise<Response> {
+    const parsed = new URL(url);
+    const isHttps = parsed.protocol === "https:";
+    const lib = isHttps ? https : http;
+    const agent =
+        isHttps && init.skipTlsVerification ? insecureHttpsAgent : undefined;
+
+    return new Promise((resolve, reject) => {
+        if (init.signal?.aborted) {
+            reject(init.signal.reason ?? new Error("Request aborted"));
+            return;
+        }
+
+        const req = lib.request(
+            url,
+            {
+                method: init.method,
+                headers: init.headers,
+                agent
+            },
+            (res) => {
+                const headers = new Headers();
+                for (const [key, value] of Object.entries(res.headers)) {
+                    if (value === undefined) {
+                        continue;
+                    }
+                    if (Array.isArray(value)) {
+                        for (const entry of value) {
+                            headers.append(key, entry);
+                        }
+                    } else {
+                        headers.set(key, value);
+                    }
+                }
+
+                const body = Readable.toWeb(res) as ReadableStream<Uint8Array>;
+                resolve(
+                    new Response(body, {
+                        status: res.statusCode ?? 502,
+                        statusText: res.statusMessage,
+                        headers
+                    })
+                );
+            }
+        );
+
+        req.on("error", reject);
+
+        if (init.signal) {
+            const onAbort = () => req.destroy(init.signal!.reason);
+            init.signal.addEventListener("abort", onAbort, { once: true });
+            req.on("close", () =>
+                init.signal!.removeEventListener("abort", onAbort)
+            );
+        }
+
+        if (init.body !== undefined) {
+            req.write(init.body);
+        }
+        req.end();
+    });
+}
@@ -0,0 +1,734 @@
+import { and, eq, inArray } from "drizzle-orm";
+import {
+    aiModels,
+    aiProviders,
+    db,
+    resourceAiModels,
+    resourceAiProviders,
+    siteResourceAiModels,
+    siteResourceAiProviders,
+    type Transaction
+} from "@server/db";
+import { z } from "zod";
+import { modelKeysConflict } from "@server/lib/aiModelKeyMatch";
+
+type DbOrTrx = Transaction | typeof db;
+
+export const modelListTypeSchema = z.enum(["allow", "block"]);
+
+export type ModelListType = z.infer<typeof modelListTypeSchema>;
+
+export const accessModeSchema = z.enum(["inherit", "select"]);
+
+export type AccessMode = z.infer<typeof accessModeSchema>;
+
+export const resourceAiProviderAttachmentSchema = z.strictObject({
+    providerId: z.number().int().positive(),
+    accessMode: accessModeSchema.optional().default("inherit"),
+    enabled: z.boolean().optional().default(true)
+});
+
+export type ResourceAiProviderInput = z.infer<
+    typeof resourceAiProviderAttachmentSchema
+>;
+
+export type ResourceAiProviderAttachment = {
+    providerId: number;
+    accessMode: AccessMode;
+    enabled: boolean;
+};
+
+export const resourceAiModelEntrySchema = z.strictObject({
+    modelId: z.number().int().positive(),
+    listType: modelListTypeSchema
+});
+
+export type ResourceAiModelEntry = z.infer<typeof resourceAiModelEntrySchema>;
+
+export type InferenceFieldsError = {
+    error: string;
+};
+
+export function isInferenceFieldsError(
+    value: { error: string } | object
+): value is InferenceFieldsError {
+    return "error" in value;
+}
+
+/**
+ * Resolve which allow/block patterns apply for an attachment.
+ * inherit → provider lists; select → resource-selected lists (replace).
+ */
+export function resolveEffectiveLists(input: {
+    accessMode: AccessMode;
+    providerAllows: string[];
+    providerBlocks: string[];
+    resourceAllows: string[];
+    resourceBlocks: string[];
+}): { allows: string[]; blocks: string[] } {
+    if (input.accessMode === "select") {
+        return {
+            allows: input.resourceAllows,
+            blocks: input.resourceBlocks
+        };
+    }
+    return {
+        allows: input.providerAllows,
+        blocks: input.providerBlocks
+    };
+}
+
+function normalizeAttachments(
+    inputs: ResourceAiProviderInput[]
+): ResourceAiProviderAttachment[] {
+    const byProviderId = new Map<
+        number,
+        { accessMode: AccessMode; enabled: boolean }
+    >();
+    for (const input of inputs) {
+        byProviderId.set(input.providerId, {
+            accessMode: input.accessMode ?? "inherit",
+            enabled: input.enabled ?? true
+        });
+    }
+    return [...byProviderId.entries()].map(
+        ([providerId, { accessMode, enabled }]) => ({
+            providerId,
+            accessMode,
+            enabled
+        })
+    );
+}
+
+type EffectiveAllowRow = {
+    providerId: number;
+    modelKey: string;
+};
+
+/**
+ * Ensure effective allow modelKeys do not conflict across attached providers.
+ * inherit uses provider allows; select uses resource-selected allows (or the
+ * optional override map). Block patterns are ignored for overlap checks.
+ */
+export async function assertNoOverlappingModelKeys(
+    attachments: ResourceAiProviderAttachment[],
+    options: {
+        trx?: DbOrTrx;
+        resourceId?: number;
+        siteResourceId?: number;
+        selectedAllowsByProvider?: Map<number, string[]>;
+    } = {}
+): Promise<InferenceFieldsError | null> {
+    const trx = options.trx ?? db;
+
+    const activeAttachments = attachments.filter((a) => a.enabled);
+
+    if (activeAttachments.length < 2) {
+        return null;
+    }
+
+    const inheritProviderIds = activeAttachments
+        .filter((a) => a.accessMode === "inherit")
+        .map((a) => a.providerId);
+    const selectProviderIds = activeAttachments
+        .filter((a) => a.accessMode === "select")
+        .map((a) => a.providerId);
+
+    const effectiveAllows: EffectiveAllowRow[] = [];
+
+    if (inheritProviderIds.length > 0) {
+        const providerAllows = await trx
+            .select({
+                providerId: aiModels.providerId,
+                modelKey: aiModels.modelKey
+            })
+            .from(aiModels)
+            .where(
+                and(
+                    inArray(aiModels.providerId, inheritProviderIds),
+                    eq(aiModels.enabled, true),
+                    eq(aiModels.listType, "allow")
+                )
+            );
+        effectiveAllows.push(...providerAllows);
+    }
+
+    if (selectProviderIds.length > 0) {
+        if (options.selectedAllowsByProvider) {
+            for (const providerId of selectProviderIds) {
+                const keys =
+                    options.selectedAllowsByProvider.get(providerId) ?? [];
+                for (const modelKey of keys) {
+                    effectiveAllows.push({ providerId, modelKey });
+                }
+            }
+        } else if (options.resourceId !== undefined) {
+            const rows = await trx
+                .select({
+                    providerId: aiModels.providerId,
+                    modelKey: aiModels.modelKey
+                })
+                .from(resourceAiModels)
+                .innerJoin(
+                    aiModels,
+                    eq(resourceAiModels.modelId, aiModels.modelId)
+                )
+                .where(
+                    and(
+                        eq(resourceAiModels.resourceId, options.resourceId),
+                        inArray(aiModels.providerId, selectProviderIds),
+                        eq(resourceAiModels.listType, "allow"),
+                        eq(aiModels.enabled, true)
+                    )
+                );
+            effectiveAllows.push(...rows);
+        } else if (options.siteResourceId !== undefined) {
+            const rows = await trx
+                .select({
+                    providerId: aiModels.providerId,
+                    modelKey: aiModels.modelKey
+                })
+                .from(siteResourceAiModels)
+                .innerJoin(
+                    aiModels,
+                    eq(siteResourceAiModels.modelId, aiModels.modelId)
+                )
+                .where(
+                    and(
+                        eq(
+                            siteResourceAiModels.siteResourceId,
+                            options.siteResourceId
+                        ),
+                        inArray(aiModels.providerId, selectProviderIds),
+                        eq(siteResourceAiModels.listType, "allow"),
+                        eq(aiModels.enabled, true)
+                    )
+                );
+            effectiveAllows.push(...rows);
+        }
+    }
+
+    const conflictPairs: string[] = [];
+    for (let i = 0; i < effectiveAllows.length; i++) {
+        for (let j = i + 1; j < effectiveAllows.length; j++) {
+            const left = effectiveAllows[i];
+            const right = effectiveAllows[j];
+            if (left.providerId === right.providerId) {
+                continue;
+            }
+            if (!modelKeysConflict(left.modelKey, right.modelKey)) {
+                continue;
+            }
+            const pair = [left.modelKey, right.modelKey].sort().join(" vs ");
+            if (!conflictPairs.includes(pair)) {
+                conflictPairs.push(pair);
+            }
+        }
+    }
+
+    if (conflictPairs.length === 0) {
+        return null;
+    }
+
+    conflictPairs.sort();
+    return {
+        error: `Model keys must be unique across providers on a resource. Overlapping keys: ${conflictPairs.join(", ")}`
+    };
+}
+
+/**
+ * Validate provider attachments for an org.
+ */
+export async function resolveProviderAttachments(input: {
+    orgId: string;
+    attachments: ResourceAiProviderInput[];
+    requireAtLeastOne: boolean;
+    resourceId?: number;
+    siteResourceId?: number;
+}): Promise<ResourceAiProviderAttachment[] | InferenceFieldsError> {
+    const attachments = normalizeAttachments(input.attachments);
+
+    if (input.requireAtLeastOne && attachments.length === 0) {
+        return {
+            error: "At least one AI provider is required for inference-mode resources"
+        };
+    }
+
+    if (attachments.length === 0) {
+        return [];
+    }
+
+    const providerIds = attachments.map((a) => a.providerId);
+    const providers = await db
+        .select({
+            providerId: aiProviders.providerId,
+            orgId: aiProviders.orgId,
+            enabled: aiProviders.enabled
+        })
+        .from(aiProviders)
+        .where(
+            and(
+                inArray(aiProviders.providerId, providerIds),
+                eq(aiProviders.orgId, input.orgId)
+            )
+        );
+
+    if (providers.length !== providerIds.length) {
+        return {
+            error: "One or more AI providers were not found in this organization"
+        };
+    }
+
+    const disabled = providers.find((p) => !p.enabled);
+    if (disabled) {
+        return {
+            error: `AI provider with ID ${disabled.providerId} is disabled`
+        };
+    }
+
+    const overlapError = await assertNoOverlappingModelKeys(attachments, {
+        resourceId: input.resourceId,
+        siteResourceId: input.siteResourceId
+    });
+    if (overlapError) {
+        return overlapError;
+    }
+
+    return attachments;
+}
+
+export async function assertInferenceModeAllowsProviderFields(input: {
+    mode: string;
+    hasProviderAttachments: boolean;
+}): Promise<InferenceFieldsError | null> {
+    if (input.mode === "inference") {
+        return null;
+    }
+    if (input.hasProviderAttachments) {
+        return {
+            error: "AI providers can only be attached to inference-mode resources"
+        };
+    }
+    return null;
+}
+
+/**
+ * Attach providers to a resource. Inherit attachments use the provider lists
+ * as-is (resource model rows for those providers are pruned). Select
+ * attachments keep resource-selected allow/block subsets.
+ */
+export async function setPublicResourceAiProviders(
+    resourceId: number,
+    attachments: ResourceAiProviderAttachment[],
+    trx: DbOrTrx = db
+): Promise<void> {
+    await trx
+        .delete(resourceAiProviders)
+        .where(eq(resourceAiProviders.resourceId, resourceId));
+
+    if (attachments.length > 0) {
+        await trx.insert(resourceAiProviders).values(
+            attachments.map((a) => ({
+                resourceId,
+                providerId: a.providerId,
+                accessMode: a.accessMode,
+                enabled: a.enabled
+            }))
+        );
+    }
+
+    await prunePublicResourceModelsToSelectProviders(
+        resourceId,
+        attachments,
+        trx
+    );
+}
+
+export async function setSiteResourceAiProviders(
+    siteResourceId: number,
+    attachments: ResourceAiProviderAttachment[],
+    trx: DbOrTrx = db
+): Promise<void> {
+    await trx
+        .delete(siteResourceAiProviders)
+        .where(eq(siteResourceAiProviders.siteResourceId, siteResourceId));
+
+    if (attachments.length > 0) {
+        await trx.insert(siteResourceAiProviders).values(
+            attachments.map((a) => ({
+                siteResourceId,
+                providerId: a.providerId,
+                accessMode: a.accessMode,
+                enabled: a.enabled
+            }))
+        );
+    }
+
+    await pruneSiteResourceModelsToSelectProviders(
+        siteResourceId,
+        attachments,
+        trx
+    );
+}
+
+/**
+ * Keep resource model rows only for providers in select mode.
+ */
+async function prunePublicResourceModelsToSelectProviders(
+    resourceId: number,
+    attachments: ResourceAiProviderAttachment[],
+    trx: DbOrTrx
+): Promise<void> {
+    const selectProviderIds = attachments
+        .filter((a) => a.accessMode === "select")
+        .map((a) => a.providerId);
+
+    if (selectProviderIds.length === 0) {
+        await trx
+            .delete(resourceAiModels)
+            .where(eq(resourceAiModels.resourceId, resourceId));
+        return;
+    }
+
+    const existing = await trx
+        .select({
+            modelId: resourceAiModels.modelId,
+            providerId: aiModels.providerId
+        })
+        .from(resourceAiModels)
+        .innerJoin(aiModels, eq(resourceAiModels.modelId, aiModels.modelId))
+        .where(eq(resourceAiModels.resourceId, resourceId));
+
+    const allowed = new Set(selectProviderIds);
+    const toRemove = existing
+        .filter((row) => !allowed.has(row.providerId))
+        .map((row) => row.modelId);
+
+    if (toRemove.length > 0) {
+        await trx
+            .delete(resourceAiModels)
+            .where(
+                and(
+                    eq(resourceAiModels.resourceId, resourceId),
+                    inArray(resourceAiModels.modelId, toRemove)
+                )
+            );
+    }
+}
+
+async function pruneSiteResourceModelsToSelectProviders(
+    siteResourceId: number,
+    attachments: ResourceAiProviderAttachment[],
+    trx: DbOrTrx
+): Promise<void> {
+    const selectProviderIds = attachments
+        .filter((a) => a.accessMode === "select")
+        .map((a) => a.providerId);
+
+    if (selectProviderIds.length === 0) {
+        await trx
+            .delete(siteResourceAiModels)
+            .where(eq(siteResourceAiModels.siteResourceId, siteResourceId));
+        return;
+    }
+
+    const existing = await trx
+        .select({
+            modelId: siteResourceAiModels.modelId,
+            providerId: aiModels.providerId
+        })
+        .from(siteResourceAiModels)
+        .innerJoin(aiModels, eq(siteResourceAiModels.modelId, aiModels.modelId))
+        .where(eq(siteResourceAiModels.siteResourceId, siteResourceId));
+
+    const allowed = new Set(selectProviderIds);
+    const toRemove = existing
+        .filter((row) => !allowed.has(row.providerId))
+        .map((row) => row.modelId);
+
+    if (toRemove.length > 0) {
+        await trx
+            .delete(siteResourceAiModels)
+            .where(
+                and(
+                    eq(siteResourceAiModels.siteResourceId, siteResourceId),
+                    inArray(siteResourceAiModels.modelId, toRemove)
+                )
+            );
+    }
+}
+
+export async function clearPublicResourceAiConfig(
+    resourceId: number,
+    trx: DbOrTrx = db
+): Promise<void> {
+    await trx
+        .delete(resourceAiModels)
+        .where(eq(resourceAiModels.resourceId, resourceId));
+    await trx
+        .delete(resourceAiProviders)
+        .where(eq(resourceAiProviders.resourceId, resourceId));
+}
+
+export async function clearSiteResourceAiConfig(
+    siteResourceId: number,
+    trx: DbOrTrx = db
+): Promise<void> {
+    await trx
+        .delete(siteResourceAiModels)
+        .where(eq(siteResourceAiModels.siteResourceId, siteResourceId));
+    await trx
+        .delete(siteResourceAiProviders)
+        .where(eq(siteResourceAiProviders.siteResourceId, siteResourceId));
+}
+
+export async function listPublicResourceAiProviders(resourceId: number) {
+    return db
+        .select({
+            providerId: resourceAiProviders.providerId,
+            name: aiProviders.name,
+            type: aiProviders.type,
+            enabled: resourceAiProviders.enabled,
+            providerEnabled: aiProviders.enabled,
+            accessMode: resourceAiProviders.accessMode
+        })
+        .from(resourceAiProviders)
+        .innerJoin(
+            aiProviders,
+            eq(resourceAiProviders.providerId, aiProviders.providerId)
+        )
+        .where(eq(resourceAiProviders.resourceId, resourceId));
+}
+
+export async function listSiteResourceAiProviders(siteResourceId: number) {
+    return db
+        .select({
+            providerId: siteResourceAiProviders.providerId,
+            name: aiProviders.name,
+            type: aiProviders.type,
+            enabled: siteResourceAiProviders.enabled,
+            providerEnabled: aiProviders.enabled,
+            accessMode: siteResourceAiProviders.accessMode
+        })
+        .from(siteResourceAiProviders)
+        .innerJoin(
+            aiProviders,
+            eq(siteResourceAiProviders.providerId, aiProviders.providerId)
+        )
+        .where(eq(siteResourceAiProviders.siteResourceId, siteResourceId));
+}
+
+/**
+ * Model list APIs require an inference resource with at least one select-mode
+ * attached provider.
+ */
+export async function assertPublicModelListApiEligible(resource: {
+    resourceId: number;
+    mode: string;
+}): Promise<string | null> {
+    if (resource.mode !== "inference") {
+        return "AI model lists are only supported on inference-mode resources";
+    }
+
+    const [row] = await db
+        .select({ providerId: resourceAiProviders.providerId })
+        .from(resourceAiProviders)
+        .where(
+            and(
+                eq(resourceAiProviders.resourceId, resource.resourceId),
+                eq(resourceAiProviders.accessMode, "select")
+            )
+        )
+        .limit(1);
+
+    if (!row) {
+        return "Set at least one attached AI provider to select mode before managing model lists";
+    }
+    return null;
+}
+
+export async function assertSiteModelListApiEligible(siteResource: {
+    siteResourceId: number;
+    mode: string;
+}): Promise<string | null> {
+    if (siteResource.mode !== "inference") {
+        return "AI model lists are only supported on inference-mode resources";
+    }
+
+    const [row] = await db
+        .select({ providerId: siteResourceAiProviders.providerId })
+        .from(siteResourceAiProviders)
+        .where(
+            and(
+                eq(
+                    siteResourceAiProviders.siteResourceId,
+                    siteResource.siteResourceId
+                ),
+                eq(siteResourceAiProviders.accessMode, "select")
+            )
+        )
+        .limit(1);
+
+    if (!row) {
+        return "Set at least one attached AI provider to select mode before managing model lists";
+    }
+    return null;
+}
+
+/**
+ * Resource model entries must belong to select-mode attached providers, and
+ * listType must match the provider catalog entry (allow→allow, block→block).
+ */
+export async function assertPublicResourceModelEntriesValid(input: {
+    orgId: string;
+    resourceId: number;
+    models: ResourceAiModelEntry[];
+}): Promise<string | null> {
+    const uniqueModels = dedupeModelEntries(input.models);
+    if (uniqueModels.length === 0) {
+        return null;
+    }
+
+    const attachments = await db
+        .select({
+            providerId: resourceAiProviders.providerId,
+            accessMode: resourceAiProviders.accessMode,
+            enabled: resourceAiProviders.enabled
+        })
+        .from(resourceAiProviders)
+        .innerJoin(
+            aiProviders,
+            eq(resourceAiProviders.providerId, aiProviders.providerId)
+        )
+        .where(
+            and(
+                eq(resourceAiProviders.resourceId, input.resourceId),
+                eq(aiProviders.orgId, input.orgId)
+            )
+        );
+
+    return assertModelEntriesValid({
+        orgId: input.orgId,
+        modelEntries: uniqueModels,
+        attachments,
+        resourceLabel: "resource"
+    });
+}
+
+export async function assertSiteResourceModelEntriesValid(input: {
+    orgId: string;
+    siteResourceId: number;
+    models: ResourceAiModelEntry[];
+}): Promise<string | null> {
+    const uniqueModels = dedupeModelEntries(input.models);
+    if (uniqueModels.length === 0) {
+        return null;
+    }
+
+    const attachments = await db
+        .select({
+            providerId: siteResourceAiProviders.providerId,
+            accessMode: siteResourceAiProviders.accessMode,
+            enabled: siteResourceAiProviders.enabled
+        })
+        .from(siteResourceAiProviders)
+        .innerJoin(
+            aiProviders,
+            eq(siteResourceAiProviders.providerId, aiProviders.providerId)
+        )
+        .where(
+            and(
+                eq(
+                    siteResourceAiProviders.siteResourceId,
+                    input.siteResourceId
+                ),
+                eq(aiProviders.orgId, input.orgId)
+            )
+        );
+
+    return assertModelEntriesValid({
+        orgId: input.orgId,
+        modelEntries: uniqueModels,
+        attachments,
+        resourceLabel: "site resource"
+    });
+}
+
+function dedupeModelEntries(
+    models: ResourceAiModelEntry[]
+): ResourceAiModelEntry[] {
+    const byModelId = new Map(
+        models.map((m) => [m.modelId, m.listType] as const)
+    );
+    return [...byModelId.entries()].map(([modelId, listType]) => ({
+        modelId,
+        listType
+    }));
+}
+
+async function assertModelEntriesValid(input: {
+    orgId: string;
+    modelEntries: ResourceAiModelEntry[];
+    attachments: ResourceAiProviderAttachment[];
+    resourceLabel: string;
+}): Promise<string | null> {
+    const selectProviderIds = input.attachments
+        .filter((a) => a.accessMode === "select")
+        .map((a) => a.providerId);
+
+    if (selectProviderIds.length === 0) {
+        return "Set at least one attached AI provider to select mode before managing model lists";
+    }
+
+    const modelIds = input.modelEntries.map((m) => m.modelId);
+    const catalogRows = await db
+        .select({
+            modelId: aiModels.modelId,
+            modelKey: aiModels.modelKey,
+            listType: aiModels.listType,
+            providerId: aiModels.providerId,
+            enabled: aiModels.enabled
+        })
+        .from(aiModels)
+        .innerJoin(aiProviders, eq(aiModels.providerId, aiProviders.providerId))
+        .where(
+            and(
+                inArray(aiModels.modelId, modelIds),
+                inArray(aiModels.providerId, selectProviderIds),
+                eq(aiProviders.orgId, input.orgId)
+            )
+        );
+
+    if (catalogRows.length !== modelIds.length) {
+        return `One or more model IDs do not exist or do not belong to a select-mode provider on this ${input.resourceLabel}`;
+    }
+
+    const catalogById = new Map(catalogRows.map((row) => [row.modelId, row]));
+    const selectedAllowsByProvider = new Map<number, string[]>();
+    for (const entry of input.modelEntries) {
+        const catalog = catalogById.get(entry.modelId);
+        if (!catalog) {
+            return `One or more model IDs do not exist or do not belong to a select-mode provider on this ${input.resourceLabel}`;
+        }
+        if (catalog.listType !== entry.listType) {
+            return `Model ${entry.modelId} must use listType "${catalog.listType}" to match the provider catalog entry`;
+        }
+        if (!catalog.enabled) {
+            return `Model ${entry.modelId} is disabled on its provider`;
+        }
+        if (entry.listType === "allow") {
+            const keys = selectedAllowsByProvider.get(catalog.providerId) ?? [];
+            keys.push(catalog.modelKey);
+            selectedAllowsByProvider.set(catalog.providerId, keys);
+        }
+    }
+
+    const overlapError = await assertNoOverlappingModelKeys(input.attachments, {
+        selectedAllowsByProvider
+    });
+    if (overlapError) {
+        return overlapError.error;
+    }
+
+    return null;
+}
@@ -0,0 +1,124 @@
+const modelKeyRegexCache = new Map<string, RegExp>();
+
+export function isModelKeyPattern(key: string): boolean {
+    return key.includes("*") || key.includes("?");
+}
+
+function getModelKeyRegex(pattern: string): RegExp {
+    let regex = modelKeyRegexCache.get(pattern);
+    if (!regex) {
+        const escaped = pattern.replace(/[.+^${}()|[\]\\]/g, "\\$&");
+        regex = new RegExp(
+            `^${escaped.replace(/\*/g, ".*").replace(/\?/g, ".")}$`
+        );
+        modelKeyRegexCache.set(pattern, regex);
+    }
+    return regex;
+}
+
+export function modelKeyMatches(
+    pattern: string,
+    requestedModel: string
+): boolean {
+    return getModelKeyRegex(pattern).test(requestedModel);
+}
+
+function wildcardCharCount(key: string): number {
+    let count = 0;
+    for (const char of key) {
+        if (char === "*" || char === "?") {
+            count += 1;
+        }
+    }
+    return count;
+}
+
+function literalLength(key: string): number {
+    return key.replace(/[*?]/g, "").length;
+}
+
+/**
+ * Sort comparator: more specific patterns sort before less specific ones
+ * (negative when `a` is more specific than `b`).
+ *
+ * 1. Exact keys beat patterns
+ * 2. Fewer wildcard characters win
+ * 3. Longer literal length wins
+ */
+export function compareModelKeySpecificity(a: string, b: string): number {
+    const aIsPattern = isModelKeyPattern(a);
+    const bIsPattern = isModelKeyPattern(b);
+
+    if (aIsPattern !== bIsPattern) {
+        return aIsPattern ? 1 : -1;
+    }
+
+    const wildcardDiff = wildcardCharCount(a) - wildcardCharCount(b);
+    if (wildcardDiff !== 0) {
+        return wildcardDiff;
+    }
+
+    return literalLength(b) - literalLength(a);
+}
+
+/**
+ * Attach-time conflict check. Detects identical keys and exact-vs-pattern
+ * matches. Does not attempt full glob intersection.
+ */
+export function modelKeysConflict(a: string, b: string): boolean {
+    if (a === b) {
+        return true;
+    }
+
+    const aIsPattern = isModelKeyPattern(a);
+    const bIsPattern = isModelKeyPattern(b);
+
+    if (aIsPattern === bIsPattern) {
+        return false;
+    }
+
+    if (aIsPattern) {
+        return modelKeyMatches(a, b);
+    }
+
+    return modelKeyMatches(b, a);
+}
+
+/**
+ * Provider-layer policy: empty allowlist denies all. Blocklist only applies
+ * after an allow match.
+ */
+export function isAllowedByLists(
+    requested: string,
+    allows: string[],
+    blocks: string[]
+): boolean {
+    if (allows.length === 0) {
+        return false;
+    }
+    if (!allows.some((pattern) => modelKeyMatches(pattern, requested))) {
+        return false;
+    }
+    if (blocks.some((pattern) => modelKeyMatches(pattern, requested))) {
+        return false;
+    }
+    return true;
+}
+
+/**
+ * Among allow patterns that match `requested`, return the most specific one,
+ * or null if none match.
+ */
+export function mostSpecificMatchingAllow(
+    requested: string,
+    allows: string[]
+): string | null {
+    const matching = allows.filter((pattern) =>
+        modelKeyMatches(pattern, requested)
+    );
+    if (matching.length === 0) {
+        return null;
+    }
+    matching.sort(compareModelKeySpecificity);
+    return matching[0];
+}
@@ -0,0 +1,230 @@
+import fs from "node:fs";
+import path from "node:path";
+import { APP_PATH } from "@server/lib/consts";
+import type { AiProviderType } from "@server/lib/aiProviderDefaults";
+import type { AiUsage } from "@server/lib/aiUsageExtraction";
+import logger from "@server/logger";
+
+// config/models.json is a runtime asset (same category as config.yml or the
+// MaxMind DBs) - not part of the source tree. Its shape mirrors litellm's
+// public model_prices_and_context_window.json: a flat list of
+// { id, name, provider, input_cost_per_token, output_cost_per_token,
+// cache_read_input_token_cost, output_cost_per_reasoning_token }, where
+// `provider` is litellm's provider bucket, not our AiProviderType.
+const MODELS_JSON_PATH = path.join(APP_PATH, "models.json");
+
+export type AiModelPricingEntry = {
+    id: string;
+    name: string;
+    provider: string;
+    input_cost_per_token: number | null;
+    output_cost_per_token: number | null;
+    cache_read_input_token_cost: number | null;
+    output_cost_per_reasoning_token: number | null;
+};
+
+export type AiModelPricing = {
+    inputCostPerToken: number | null;
+    outputCostPerToken: number | null;
+    cacheReadInputTokenCost: number | null;
+    outputCostPerReasoningToken: number | null;
+    // True when the match came from a different provider bucket than the one
+    // mapped to this provider's type (e.g. an openRouter/custom model id that
+    // only matched by stripping a "vendor/" prefix against the whole table).
+    // Costs found this way are a best-effort approximation, not a guarantee
+    // the upstream provider bills at the same rate.
+    approximate: boolean;
+};
+
+// Which litellm provider buckets to search for each of our provider types.
+// Several of our provider types (openRouter, vercelAiGateway, custom) proxy
+// arbitrary underlying models and have no dedicated bucket in the pricing
+// data, so they fall back to a global search across all buckets.
+const PROVIDER_PRICING_BUCKETS: Record<
+    Exclude<AiProviderType, "custom">,
+    string[]
+> = {
+    openai: ["openai"],
+    anthropic: ["anthropic"],
+    googleGemini: ["gemini"],
+    vertexAi: [
+        "vertex_ai-language-models",
+        "vertex_ai",
+        "vertex_ai-anthropic_models",
+        "vertex_ai-mistral_models",
+        "vertex_ai-deepseek_models",
+        "vertex_ai-ai21_models",
+        "vertex_ai-llama_models",
+        "vertex_ai-minimax_models",
+        "vertex_ai-moonshot_models",
+        "vertex_ai-zai_models",
+        "vertex_ai-openai_models",
+        "vertex_ai-qwen_models",
+        "vertex_ai-text-models"
+    ],
+    bedrock: ["bedrock_converse", "bedrock", "bedrock_mantle"],
+    microsoftFoundry: ["azure", "azure_ai", "azure_text"],
+    openRouter: [],
+    vercelAiGateway: []
+};
+
+let modelsById: Map<string, AiModelPricingEntry[]> | null = null;
+
+function loadModels(): Map<string, AiModelPricingEntry[]> {
+    if (modelsById) {
+        return modelsById;
+    }
+
+    const byId = new Map<string, AiModelPricingEntry[]>();
+    try {
+        if (fs.existsSync(MODELS_JSON_PATH)) {
+            const raw = fs.readFileSync(MODELS_JSON_PATH, "utf-8");
+            const parsed = JSON.parse(raw) as { data: AiModelPricingEntry[] };
+            for (const entry of parsed.data ?? []) {
+                for (const key of [entry.id, entry.name]) {
+                    if (!key) continue;
+                    const list = byId.get(key) ?? [];
+                    list.push(entry);
+                    byId.set(key, list);
+                }
+            }
+        } else {
+            logger.debug(
+                `AI model pricing file not found at ${MODELS_JSON_PATH}; cost calculation will fall back to unknown pricing`
+            );
+        }
+    } catch (error) {
+        logger.warn("Failed to load AI model pricing file", { error });
+    }
+
+    modelsById = byId;
+    return byId;
+}
+
+function stripVendorPrefix(modelId: string): string | null {
+    const idx = modelId.indexOf("/");
+    if (idx === -1 || idx === modelId.length - 1) {
+        return null;
+    }
+    return modelId.slice(idx + 1);
+}
+
+function toPricing(
+    entry: AiModelPricingEntry,
+    approximate: boolean
+): AiModelPricing {
+    return {
+        inputCostPerToken: entry.input_cost_per_token,
+        outputCostPerToken: entry.output_cost_per_token,
+        cacheReadInputTokenCost: entry.cache_read_input_token_cost,
+        outputCostPerReasoningToken: entry.output_cost_per_reasoning_token,
+        approximate
+    };
+}
+
+function findInBuckets(
+    byId: Map<string, AiModelPricingEntry[]>,
+    modelId: string,
+    buckets: string[] | null
+): AiModelPricingEntry | null {
+    const candidates = [modelId, stripVendorPrefix(modelId)].filter(
+        (v): v is string => v != null
+    );
+
+    for (const key of candidates) {
+        const entries = byId.get(key);
+        if (!entries) continue;
+        const match = buckets
+            ? entries.find((e) => buckets.includes(e.provider))
+            : entries[0];
+        if (match) {
+            return match;
+        }
+    }
+    return null;
+}
+
+/**
+ * Looks up per-token pricing for a model, scoped first to the litellm
+ * provider bucket(s) that correspond to our provider type, then falling
+ * back to a global search across all buckets (marked `approximate`) for
+ * provider types that proxy arbitrary underlying models.
+ */
+export function getModelPricing(
+    providerType: AiProviderType,
+    modelId: string | undefined
+): AiModelPricing | null {
+    if (!modelId) {
+        return null;
+    }
+
+    const byId = loadModels();
+    const buckets =
+        providerType === "custom"
+            ? []
+            : PROVIDER_PRICING_BUCKETS[providerType];
+
+    if (buckets && buckets.length > 0) {
+        const scoped = findInBuckets(byId, modelId, buckets);
+        if (scoped) {
+            return toPricing(scoped, false);
+        }
+    }
+
+    const fallback = findInBuckets(byId, modelId, null);
+    if (fallback) {
+        return toPricing(fallback, true);
+    }
+
+    return null;
+}
+
+export type AiCostBreakdown = {
+    promptCost: number;
+    cacheReadCost: number;
+    cacheWriteCost: number;
+    completionCost: number;
+    reasoningCost: number;
+    totalCost: number;
+};
+
+/**
+ * Computes a $ cost breakdown for a usage record given a model's pricing.
+ * Cache writes and reasoning tokens fall back to the normal input/output
+ * rate respectively when the pricing data has no dedicated rate for them
+ * (the models.json schema here has no cache-write field at all, and only
+ * some models report a distinct reasoning rate).
+ */
+export function calculateAiCost(
+    pricing: AiModelPricing | null,
+    usage: AiUsage
+): AiCostBreakdown | null {
+    if (!pricing) {
+        return null;
+    }
+
+    const inputRate = pricing.inputCostPerToken ?? 0;
+    const outputRate = pricing.outputCostPerToken ?? 0;
+    const cacheReadRate = pricing.cacheReadInputTokenCost ?? inputRate;
+    const reasoningRate = pricing.outputCostPerReasoningToken ?? outputRate;
+
+    const promptCost = usage.promptTokens * inputRate;
+    const cacheReadCost = usage.cacheReadTokens * cacheReadRate;
+    const cacheWriteCost = usage.cacheWriteTokens * inputRate;
+    const completionCost = usage.completionTokens * outputRate;
+    const reasoningCost = usage.reasoningTokens * reasoningRate;
+
+    return {
+        promptCost,
+        cacheReadCost,
+        cacheWriteCost,
+        completionCost,
+        reasoningCost,
+        totalCost:
+            promptCost +
+            cacheReadCost +
+            cacheWriteCost +
+            completionCost +
+            reasoningCost
+    };
+}
@@ -0,0 +1,262 @@
+import { decrypt, encrypt } from "@server/lib/crypto";
+import {
+    parseCapabilities,
+    type AiCapability
+} from "@server/lib/aiCapabilities";
+
+export type AiProviderType =
+    | "openai"
+    | "anthropic"
+    | "googleGemini"
+    | "vertexAi"
+    | "bedrock"
+    | "microsoftFoundry"
+    | "openRouter"
+    | "vercelAiGateway"
+    | "custom";
+
+export const AI_PROVIDER_AUTH_TYPES = [
+    "bearer",
+    "x-api-key",
+    "x-goog-api-key",
+    "hec",
+    "cf-aig-authorization",
+    "none",
+    "passthrough"
+] as const;
+
+export type AiProviderAuthType = (typeof AI_PROVIDER_AUTH_TYPES)[number];
+export type AiBudgetUnit = "usd" | "tokens";
+export type AiProviderRoutingMode = "url" | "target";
+
+type AiProviderDefaults = {
+    upstreamUrl: string | null;
+    authType: AiProviderAuthType;
+    capabilities: readonly AiCapability[];
+};
+
+export const AI_PROVIDER_DEFAULTS: Record<
+    Exclude<AiProviderType, "custom">,
+    AiProviderDefaults
+> = {
+    openai: {
+        upstreamUrl: "https://api.openai.com/v1",
+        authType: "bearer",
+        capabilities: ["openai_chat", "openai_responses"]
+    },
+    anthropic: {
+        upstreamUrl: "https://api.anthropic.com",
+        authType: "x-api-key",
+        capabilities: ["anthropic_messages"]
+    },
+    googleGemini: {
+        upstreamUrl: "https://generativelanguage.googleapis.com",
+        authType: "x-goog-api-key",
+        capabilities: ["gemini_generate_content"]
+    },
+    vertexAi: {
+        upstreamUrl: null,
+        authType: "bearer",
+        capabilities: ["google_generate_content", "google_raw_predict"]
+    },
+    bedrock: {
+        upstreamUrl: "https://bedrock-runtime.us-east-1.amazonaws.com",
+        authType: "bearer",
+        capabilities: ["bedrock_converse"]
+    },
+    microsoftFoundry: {
+        upstreamUrl: null,
+        authType: "bearer",
+        capabilities: ["openai_chat", "openai_responses", "anthropic_messages"]
+    },
+    openRouter: {
+        upstreamUrl: "https://openrouter.ai/api/v1",
+        authType: "bearer",
+        capabilities: ["openai_chat"]
+    },
+    vercelAiGateway: {
+        upstreamUrl: "https://ai-gateway.vercel.sh/v1",
+        authType: "bearer",
+        capabilities: ["openai_chat", "openai_responses"]
+    }
+};
+
+const CONFLICTING_AUTH_HEADERS = [
+    "authorization",
+    "x-api-key",
+    "x-goog-api-key",
+    "cf-aig-authorization"
+] as const;
+
+export function authTypeRequiresApiKey(authType: AiProviderAuthType): boolean {
+    return authType !== "none" && authType !== "passthrough";
+}
+
+export function providerRequiresUpstreamUrl(
+    type: AiProviderType,
+    routingMode: AiProviderRoutingMode = "url"
+): boolean {
+    if (routingMode === "target") {
+        return false;
+    }
+    if (type === "custom") {
+        return true;
+    }
+    return AI_PROVIDER_DEFAULTS[type].upstreamUrl === null;
+}
+
+export function resolveAiProviderCreateFields(input: {
+    type: AiProviderType;
+    upstreamUrl?: string | null;
+    authType?: AiProviderAuthType | null;
+    routingMode?: AiProviderRoutingMode | null;
+}): {
+    upstreamUrl: string | null;
+    authType: AiProviderAuthType;
+    routingMode: AiProviderRoutingMode;
+} {
+    const routingMode =
+        input.type === "custom" ? (input.routingMode ?? "url") : "url";
+
+    if (routingMode === "target") {
+        return {
+            upstreamUrl: null,
+            authType: input.authType ?? "bearer",
+            routingMode
+        };
+    }
+
+    if (input.type === "custom") {
+        return {
+            upstreamUrl: input.upstreamUrl ?? null,
+            authType: input.authType ?? "bearer",
+            routingMode
+        };
+    }
+
+    const defaults = AI_PROVIDER_DEFAULTS[input.type];
+    return {
+        upstreamUrl: input.upstreamUrl ?? defaults.upstreamUrl,
+        authType: input.authType ?? defaults.authType,
+        routingMode
+    };
+}
+
+export type AiProviderHeader = { name: string; value: string };
+
+export function serializeAiProviderHeaders(
+    headers: AiProviderHeader[] | null | undefined,
+    secret: string
+): string | null {
+    if (!headers || headers.length === 0) {
+        return null;
+    }
+    return encrypt(JSON.stringify(headers), secret);
+}
+
+export function parseAiProviderHeaders(
+    raw: string | null | undefined,
+    secret: string
+): AiProviderHeader[] {
+    if (!raw) {
+        return [];
+    }
+    try {
+        const decrypted = decrypt(raw, secret);
+        const parsed = JSON.parse(decrypted);
+        if (!Array.isArray(parsed)) {
+            return [];
+        }
+        return parsed.filter(
+            (h): h is AiProviderHeader =>
+                h != null &&
+                typeof h === "object" &&
+                typeof h.name === "string" &&
+                typeof h.value === "string"
+        );
+    } catch {
+        return [];
+    }
+}
+
+export function applyAiProviderCustomHeaders(
+    headers: Record<string, string>,
+    raw: string | null | undefined,
+    secret: string
+): void {
+    for (const { name, value } of parseAiProviderHeaders(raw, secret)) {
+        headers[name] = value;
+    }
+}
+
+/**
+ * Apply provider auth to upstream headers.
+ * - Injected modes: strip client auth headers, then set the provider key.
+ * - none: strip client auth headers, send no auth.
+ * - passthrough: leave client auth headers as-is.
+ */
+export function applyAiProviderAuthHeaders(
+    headers: Record<string, string>,
+    authType: AiProviderAuthType,
+    apiKey: string | null
+): void {
+    if (authType === "passthrough") {
+        return;
+    }
+
+    for (const name of CONFLICTING_AUTH_HEADERS) {
+        for (const key of Object.keys(headers)) {
+            if (key.toLowerCase() === name) {
+                delete headers[key];
+            }
+        }
+    }
+
+    if (authType === "none") {
+        return;
+    }
+
+    if (!apiKey) {
+        throw new Error(`API key required for authType ${authType}`);
+    }
+
+    switch (authType) {
+        case "bearer":
+            headers["Authorization"] = `Bearer ${apiKey}`;
+            break;
+        case "x-api-key":
+            headers["x-api-key"] = apiKey;
+            break;
+        case "x-goog-api-key":
+            headers["x-goog-api-key"] = apiKey;
+            break;
+        case "hec":
+            headers["Authorization"] = `Splunk ${apiKey}`;
+            break;
+        case "cf-aig-authorization":
+            headers["cf-aig-authorization"] = `Bearer ${apiKey}`;
+            break;
+    }
+}
+
+export function resolveCapabilitiesForCreate(input: {
+    type: AiProviderType;
+    capabilities?: AiCapability[] | null;
+}): AiCapability[] {
+    if (input.capabilities != null) {
+        return parseCapabilities(input.capabilities);
+    }
+    if (input.type === "custom") {
+        return [];
+    }
+    return [...AI_PROVIDER_DEFAULTS[input.type].capabilities];
+}
+
+export function defaultsForProviderType(
+    type: AiProviderType
+): readonly AiCapability[] {
+    if (type === "custom") {
+        return [];
+    }
+    return AI_PROVIDER_DEFAULTS[type].capabilities;
+}
@@ -0,0 +1,468 @@
+import { encode } from "gpt-tokenizer";
+import type { AiCapability } from "@server/lib/aiCapabilities";
+import logger from "@server/logger";
+
+export type AiUsage = {
+    // Input tokens billed at the normal input rate (i.e. NOT already
+    // covered by cacheReadTokens/cacheWriteTokens below).
+    promptTokens: number;
+    cacheReadTokens: number;
+    cacheWriteTokens: number;
+    // Output tokens billed at the normal output rate (i.e. NOT already
+    // covered by reasoningTokens below).
+    completionTokens: number;
+    reasoningTokens: number;
+    // True when these numbers are our own best-guess estimate (the upstream
+    // response didn't report usage), rather than provider-reported figures.
+    estimated: boolean;
+};
+
+function emptyUsage(): AiUsage {
+    return {
+        promptTokens: 0,
+        cacheReadTokens: 0,
+        cacheWriteTokens: 0,
+        completionTokens: 0,
+        reasoningTokens: 0,
+        estimated: false
+    };
+}
+
+/**
+ * Scans raw (possibly binary-framed, e.g. Bedrock's vnd.amazon.eventstream)
+ * text for `"fieldName":123` occurrences and returns the last value seen for
+ * each field. Used as a best-effort fallback for response shapes we can't
+ * fully parse as JSON/SSE (streaming Bedrock, raw predict passthroughs).
+ */
+function scanNumericFields(
+    text: string,
+    fields: string[]
+): Record<string, number> {
+    const out: Record<string, number> = {};
+    for (const field of fields) {
+        const re = new RegExp(`"${field}"\\s*:\\s*(\\d+)`, "g");
+        let match: RegExpExecArray | null;
+        while ((match = re.exec(text)) !== null) {
+            out[field] = Number(match[1]);
+        }
+    }
+    return out;
+}
+
+function sseDataFrames(text: string): string[] {
+    const frames: string[] = [];
+    for (const rawFrame of text.split(/\r?\n\r?\n/)) {
+        for (const line of rawFrame.split(/\r?\n/)) {
+            if (!line.startsWith("data:")) continue;
+            const data = line.slice("data:".length).trim();
+            if (data && data !== "[DONE]") {
+                frames.push(data);
+            }
+        }
+    }
+    return frames;
+}
+
+function tryParseJson(text: string): any | null {
+    try {
+        return JSON.parse(text);
+    } catch {
+        return null;
+    }
+}
+
+function extractOpenAiChat(text: string, isStream: boolean): AiUsage | null {
+    let usage: any = null;
+
+    if (isStream) {
+        for (const frame of sseDataFrames(text)) {
+            const parsed = tryParseJson(frame);
+            if (parsed?.usage) {
+                usage = parsed.usage;
+            }
+        }
+    } else {
+        usage = tryParseJson(text)?.usage ?? null;
+    }
+
+    if (!usage) {
+        return null;
+    }
+
+    const cacheReadTokens = usage.prompt_tokens_details?.cached_tokens ?? 0;
+    const reasoningTokens =
+        usage.completion_tokens_details?.reasoning_tokens ?? 0;
+
+    return {
+        promptTokens: Math.max(0, (usage.prompt_tokens ?? 0) - cacheReadTokens),
+        cacheReadTokens,
+        cacheWriteTokens: 0,
+        completionTokens: Math.max(
+            0,
+            (usage.completion_tokens ?? 0) - reasoningTokens
+        ),
+        reasoningTokens,
+        estimated: false
+    };
+}
+
+function extractOpenAiResponses(
+    text: string,
+    isStream: boolean
+): AiUsage | null {
+    let usage: any = null;
+
+    if (isStream) {
+        for (const frame of sseDataFrames(text)) {
+            const parsed = tryParseJson(frame);
+            if (parsed?.type === "response.completed" && parsed?.response?.usage) {
+                usage = parsed.response.usage;
+            } else if (parsed?.usage) {
+                usage = parsed.usage;
+            }
+        }
+    } else {
+        const parsed = tryParseJson(text);
+        usage = parsed?.usage ?? parsed?.response?.usage ?? null;
+    }
+
+    if (!usage) {
+        return null;
+    }
+
+    const cacheReadTokens = usage.input_tokens_details?.cached_tokens ?? 0;
+    const reasoningTokens = usage.output_tokens_details?.reasoning_tokens ?? 0;
+
+    return {
+        promptTokens: Math.max(0, (usage.input_tokens ?? 0) - cacheReadTokens),
+        cacheReadTokens,
+        cacheWriteTokens: 0,
+        completionTokens: Math.max(
+            0,
+            (usage.output_tokens ?? 0) - reasoningTokens
+        ),
+        reasoningTokens,
+        estimated: false
+    };
+}
+
+function extractAnthropicMessages(
+    text: string,
+    isStream: boolean
+): AiUsage | null {
+    let inputTokens = 0;
+    let cacheReadTokens = 0;
+    let cacheWriteTokens = 0;
+    let outputTokens = 0;
+    let found = false;
+
+    const applyUsage = (usage: any) => {
+        if (!usage) return;
+        found = true;
+        if (typeof usage.input_tokens === "number") {
+            inputTokens = usage.input_tokens;
+        }
+        if (typeof usage.cache_read_input_tokens === "number") {
+            cacheReadTokens = usage.cache_read_input_tokens;
+        }
+        if (typeof usage.cache_creation_input_tokens === "number") {
+            cacheWriteTokens = usage.cache_creation_input_tokens;
+        }
+        if (typeof usage.output_tokens === "number") {
+            outputTokens = usage.output_tokens;
+        }
+    };
+
+    if (isStream) {
+        for (const frame of sseDataFrames(text)) {
+            const parsed = tryParseJson(frame);
+            if (!parsed) continue;
+            applyUsage(parsed.message?.usage);
+            applyUsage(parsed.usage);
+        }
+    } else {
+        applyUsage(tryParseJson(text)?.usage);
+    }
+
+    if (!found) {
+        return null;
+    }
+
+    return {
+        promptTokens: inputTokens,
+        cacheReadTokens,
+        cacheWriteTokens,
+        completionTokens: outputTokens,
+        // Anthropic bills extended-thinking output at the normal output
+        // rate, so there's no separate reasoning bucket to report.
+        reasoningTokens: 0,
+        estimated: false
+    };
+}
+
+function extractGoogleGenerateContent(
+    text: string,
+    _isStream: boolean
+): AiUsage | null {
+    // Both the plain-JSON-array stream format and the SSE (?alt=sse) format
+    // repeat a cumulative `usageMetadata` object per chunk; the regex scan
+    // below naturally picks up the last (most complete) one either way.
+    const fields = scanNumericFields(text, [
+        "promptTokenCount",
+        "candidatesTokenCount",
+        "cachedContentTokenCount",
+        "thoughtsTokenCount"
+    ]);
+
+    if (fields.promptTokenCount === undefined) {
+        return null;
+    }
+
+    const cacheReadTokens = fields.cachedContentTokenCount ?? 0;
+    const reasoningTokens = fields.thoughtsTokenCount ?? 0;
+
+    return {
+        promptTokens: Math.max(0, fields.promptTokenCount - cacheReadTokens),
+        cacheReadTokens,
+        cacheWriteTokens: 0,
+        completionTokens: fields.candidatesTokenCount ?? 0,
+        reasoningTokens,
+        estimated: false
+    };
+}
+
+function extractBedrockConverse(
+    text: string,
+    _isStream: boolean
+): AiUsage | null {
+    // Non-streaming responses are plain JSON; converse-stream frames the
+    // final `metadata` event's usage object inside binary event-stream
+    // framing, but the JSON text survives intact inside that binary
+    // envelope, so the same field scan works for both.
+    const parsed = tryParseJson(text);
+    const usage = parsed?.usage;
+    if (usage) {
+        const cacheReadTokens = usage.cacheReadInputTokens ?? 0;
+        return {
+            promptTokens: Math.max(0, (usage.inputTokens ?? 0) - cacheReadTokens),
+            cacheReadTokens,
+            cacheWriteTokens: usage.cacheWriteInputTokens ?? 0,
+            completionTokens: usage.outputTokens ?? 0,
+            reasoningTokens: 0,
+            estimated: false
+        };
+    }
+
+    const fields = scanNumericFields(text, [
+        "inputTokens",
+        "outputTokens",
+        "cacheReadInputTokens",
+        "cacheWriteInputTokens"
+    ]);
+    if (fields.inputTokens === undefined) {
+        return null;
+    }
+    const cacheReadTokens = fields.cacheReadInputTokens ?? 0;
+    return {
+        promptTokens: Math.max(0, fields.inputTokens - cacheReadTokens),
+        cacheReadTokens,
+        cacheWriteTokens: fields.cacheWriteInputTokens ?? 0,
+        completionTokens: fields.outputTokens ?? 0,
+        reasoningTokens: 0,
+        estimated: false
+    };
+}
+
+function extractBedrockModelInvoke(
+    text: string,
+    _isStream: boolean,
+    headers: Headers
+): AiUsage | null {
+    // Non-streaming invoke reports counts via response headers regardless
+    // of the underlying model's payload format.
+    const headerInput = headers.get("x-amzn-bedrock-input-token-count");
+    const headerOutput = headers.get("x-amzn-bedrock-output-token-count");
+    if (headerInput !== null || headerOutput !== null) {
+        return {
+            promptTokens: Number(headerInput ?? 0),
+            cacheReadTokens: 0,
+            cacheWriteTokens: 0,
+            completionTokens: Number(headerOutput ?? 0),
+            reasoningTokens: 0,
+            estimated: false
+        };
+    }
+
+    // invoke-with-response-stream has no equivalent headers; the model's
+    // own usage shape (frequently Anthropic-style on Bedrock) is embedded
+    // inside binary event-stream framing, so fall back to a couple of
+    // known field-name shapes via regex.
+    const anthropicStyle = extractAnthropicMessages(text, true);
+    if (anthropicStyle) {
+        return anthropicStyle;
+    }
+
+    const fields = scanNumericFields(text, [
+        "inputTokenCount",
+        "outputTokenCount"
+    ]);
+    if (fields.inputTokenCount === undefined) {
+        return null;
+    }
+    return {
+        promptTokens: fields.inputTokenCount,
+        cacheReadTokens: 0,
+        cacheWriteTokens: 0,
+        completionTokens: fields.outputTokenCount ?? 0,
+        reasoningTokens: 0,
+        estimated: false
+    };
+}
+
+const EXTRACTORS: Record<
+    AiCapability,
+    (text: string, isStream: boolean, headers: Headers) => AiUsage | null
+> = {
+    openai_chat: extractOpenAiChat,
+    openai_responses: extractOpenAiResponses,
+    anthropic_messages: extractAnthropicMessages,
+    gemini_generate_content: extractGoogleGenerateContent,
+    google_generate_content: extractGoogleGenerateContent,
+    // rawPredict is a passthrough to whatever the underlying publisher
+    // model speaks (often Anthropic-shaped on Vertex); try that, then give
+    // up to the token-count estimate.
+    google_raw_predict: (text, isStream) =>
+        extractAnthropicMessages(text, isStream),
+    bedrock_model_invoke: extractBedrockModelInvoke,
+    bedrock_converse: extractBedrockConverse
+};
+
+/**
+ * Attempts to pull provider-reported token usage out of an upstream AI
+ * gateway response. Returns null if the response didn't contain (or we
+ * couldn't find) usage data, in which case callers should fall back to
+ * `estimateUsage`.
+ */
+export function extractUsage(
+    capability: AiCapability,
+    responseText: string,
+    isStream: boolean,
+    headers: Headers
+): AiUsage | null {
+    try {
+        return EXTRACTORS[capability](responseText, isStream, headers);
+    } catch (error) {
+        logger.debug("Failed to extract AI usage from response", {
+            capability,
+            error
+        });
+        return null;
+    }
+}
+
+/**
+ * Best-guess token estimate for when the provider doesn't report usage.
+ * Uses OpenAI's BPE tokenizer as a stand-in for whatever tokenizer the
+ * actual model uses - close enough for an approximate cost figure, not
+ * exact for non-OpenAI models.
+ */
+export function estimateUsage(
+    promptText: string,
+    completionText: string
+): AiUsage {
+    const usage = emptyUsage();
+    usage.estimated = true;
+    try {
+        usage.promptTokens = promptText ? encode(promptText).length : 0;
+    } catch (error) {
+        logger.debug("Failed to estimate prompt tokens", { error });
+    }
+    try {
+        usage.completionTokens = completionText
+            ? encode(completionText).length
+            : 0;
+    } catch (error) {
+        logger.debug("Failed to estimate completion tokens", { error });
+    }
+    return usage;
+}
+
+/**
+ * OpenAI's Chat Completions API only includes a `usage` field in a
+ * streaming response when the request opts in via `stream_options:
+ * {include_usage: true}` - unlike the Responses API, Anthropic, Gemini and
+ * Bedrock, which report usage in a streaming response by default. Returns
+ * whether we need to inject that option ourselves to be able to track cost.
+ */
+export function needsStreamUsageInjection(
+    capability: AiCapability,
+    body: any
+): boolean {
+    return (
+        capability === "openai_chat" &&
+        body?.stream === true &&
+        body?.stream_options?.include_usage !== true
+    );
+}
+
+/**
+ * Returns a shallow-cloned body with `stream_options.include_usage`
+ * injected, for capabilities/requests where `needsStreamUsageInjection`
+ * is true. Leaves the original body untouched.
+ */
+export function withStreamUsageOption(body: any): any {
+    return {
+        ...body,
+        stream_options: { ...body.stream_options, include_usage: true }
+    };
+}
+
+/**
+ * When we injected stream_options.include_usage ourselves (the caller
+ * didn't ask for it), OpenAI appends an extra terminal SSE frame with an
+ * empty `choices: []` array carrying only the usage data. Callers that
+ * don't expect that shape (most minimal SSE parsers assume a non-empty
+ * choices array) shouldn't see it, so it's stripped back out of the bytes
+ * forwarded to the client.
+ */
+export function stripInjectedUsageFrame(sseText: string): string {
+    const parts = sseText.split(/(\r?\n\r?\n)/);
+    let out = "";
+    for (let i = 0; i < parts.length; i += 2) {
+        const frame = parts[i];
+        const separator = parts[i + 1] ?? "";
+        const dataLine = frame
+            .split(/\r?\n/)
+            .find((line) => line.startsWith("data:"));
+        if (dataLine) {
+            const data = dataLine.slice("data:".length).trim();
+            const parsed = data !== "[DONE]" ? tryParseJson(data) : null;
+            if (parsed && Array.isArray(parsed.choices) && parsed.choices.length === 0 && parsed.usage) {
+                continue;
+            }
+        }
+        out += frame + separator;
+    }
+    return out;
+}
+
+/**
+ * Best-effort extraction of the model the upstream provider actually
+ * served, which some gateways/routers echo back and which may differ from
+ * the model the caller requested (e.g. an alias resolving to a dated
+ * snapshot). Falls back to the caller's requested model when absent.
+ */
+export function extractResponseModel(responseText: string): string | null {
+    const match = responseText.match(/"model"\s*:\s*"([^"]+)"/);
+    return match ? match[1] : null;
+}
+
+export function isUsageEmpty(usage: AiUsage): boolean {
+    return (
+        usage.promptTokens === 0 &&
+        usage.cacheReadTokens === 0 &&
+        usage.cacheWriteTokens === 0 &&
+        usage.completionTokens === 0 &&
+        usage.reasoningTokens === 0
+    );
+}
@@ -202,6 +202,10 @@ async function handleResource(
        return;
    }

+    if (!target.resourceId) {
+        return;
+    }
+
    const [resource] = await trx
        .select()
        .from(resources)
@@ -227,9 +231,7 @@ async function handleResource(

    let health = "healthy";
    const allUnknown = monitoredTargets.length === 0;
-    const allHealthy = monitoredTargets.every(
-        (t) => t.hcHealth === "healthy"
-    );
+    const allHealthy = monitoredTargets.every((t) => t.hcHealth === "healthy");
    const allUnhealthy = monitoredTargets.every(
        (t) => t.hcHealth === "unhealthy"
    );
@@ -339,19 +339,6 @@ export async function calculateUserClientsForOrgs(
                continue;
            }

-            // Get exit nodes for this org
-            const exitNodesList = await getExitNodes(orgId);
-
-            if (exitNodesList.length === 0) {
-                logger.warn(
-                    `Skipping org ${orgId} for OLM ${olm.olmId} (user ${userId}): no exit nodes found`
-                );
-                continue;
-            }
-
-            const randomExitNode =
-                exitNodesList[Math.floor(Math.random() * exitNodesList.length)];
-
            // Get next available subnet
            const { value: newSubnet, release: releaseSubnetLock } =
                await getNextAvailableClientSubnet(orgId, trx);
@@ -370,7 +357,6 @@ export async function calculateUserClientsForOrgs(
            const newClientData: InferInsertModel<typeof clients> = {
                userId,
                orgId: userOrg.orgId,
-                exitNodeId: randomExitNode.exitNodeId,
                name: olm.name || "User Client",
                subnet: updatedSubnet,
                olmId: olm.olmId,
@@ -93,6 +93,9 @@ export async function deleteOrgById(
            await trx.delete(sites).where(eq(sites.siteId, site.siteId));
        }
        for (const client of orgClients) {
+            if (client.exitNodeId && client.pubKey) {
+                await deletePeer(client.exitNodeId, client.pubKey);
+            }
            const [olm] = await trx
                .select()
                .from(olms)
@@ -64,13 +64,20 @@ export async function performDeleteResources(

    const targetsByResourceId = new Map<number, Target[]>();
    for (const target of targetsToBeRemoved) {
+        if (target.resourceId == null) {
+            continue;
+        }
        const existing = targetsByResourceId.get(target.resourceId) ?? [];
        existing.push(target);
        targetsByResourceId.set(target.resourceId, existing);
    }

    const targetIdToResourceId = new Map(
-        targetsToBeRemoved.map((target) => [target.targetId, target.resourceId])
+        targetsToBeRemoved.flatMap((target) =>
+            target.resourceId == null
+                ? []
+                : [[target.targetId, target.resourceId] as const]
+        )
    );

    const healthChecksByResourceId = new Map<number, TargetHealthCheck[]>();
@@ -1,4 +1,4 @@
-import { and, eq, inArray, sql } from "drizzle-orm";
+import { and, eq, inArray, isNotNull, sql } from "drizzle-orm";
 import {
    db,
    resources,
@@ -33,9 +33,11 @@ export async function getResourceIdsForSite(
    const rows = await trx
        .selectDistinct({ resourceId: targets.resourceId })
        .from(targets)
-        .where(eq(targets.siteId, siteId));
+        .where(and(eq(targets.siteId, siteId), isNotNull(targets.resourceId)));

-    return rows.map((row) => row.resourceId);
+    return rows
+        .map((row) => row.resourceId)
+        .filter((resourceId): resourceId is number => resourceId != null);
 }

 export async function getSiteResourceIdsForSite(
@@ -0,0 +1,40 @@
+import { db, sites, clients } from "@server/db";
+import { and, eq, count } from "drizzle-orm";
+
+// (MAX_CONNECTIONS - current_connections) / MAX_CONNECTIONS)
+// higher = more desirable
+// like saying, this node has x% of its capacity left
+export async function calculateExitNodeWeight(
+    exitNodeId: number,
+    maxConnections: number | null | undefined
+): Promise<number | null> {
+    if (maxConnections === null || maxConnections === undefined) {
+        return 1;
+    }
+
+    const [[siteConnections], [clientConnections]] = await Promise.all([
+        db
+            .select({ count: count() })
+            .from(sites)
+            .where(
+                and(eq(sites.exitNodeId, exitNodeId), eq(sites.online, true))
+            ),
+        db
+            .select({ count: count() })
+            .from(clients)
+            .where(
+                and(
+                    eq(clients.exitNodeId, exitNodeId),
+                    eq(clients.online, true)
+                )
+            )
+    ]);
+
+    const currentConnections = siteConnections.count + clientConnections.count;
+
+    if (currentConnections >= maxConnections) {
+        return null;
+    }
+
+    return (maxConnections - currentConnections) / maxConnections;
+}
@@ -1,6 +1,5 @@
 import { db, exitNodes, Transaction } from "@server/db";
 import logger from "@server/logger";
-import { ExitNodePingResult } from "@server/routers/newt";
 import { eq } from "drizzle-orm";

 export async function verifyExitNodeOrgAccess(
@@ -52,6 +51,16 @@ export async function listExitNodes(
    return allExitNodes;
 }

+export type ExitNodePingResult = {
+    exitNodeId: number;
+    latencyMs: number;
+    weight: number;
+    error?: string;
+    exitNodeName: string;
+    endpoint: string;
+    wasPreviouslyConnected: boolean;
+};
+
 export function selectBestExitNode(
    pingResults: ExitNodePingResult[]
 ): ExitNodePingResult | null {
@@ -0,0 +1,41 @@
+import { db, ExitNode, Transaction, sites, clients } from "@server/db";
+import { eq } from "drizzle-orm";
+import config from "@server/lib/config";
+import { findNextAvailableCidr } from "@server/lib/ip";
+import { lockManager } from "#dynamic/lib/lock";
+
+export async function getUniqueSubnetForExitNode(
+    exitNode: ExitNode,
+    trx: Transaction | typeof db = db
+): Promise<string | null> {
+    const lockKey = `subnet-allocation:${exitNode.exitNodeId}`;
+
+    return await lockManager.withLock(
+        lockKey,
+        async () => {
+            const [sitesQuery, clientsQuery] = await Promise.all([
+                trx
+                    .select({ subnet: sites.exitNodeSubnet })
+                    .from(sites)
+                    .where(eq(sites.exitNodeId, exitNode.exitNodeId)),
+                trx
+                    .select({ subnet: clients.exitNodeSubnet })
+                    .from(clients)
+                    .where(eq(clients.exitNodeId, exitNode.exitNodeId))
+            ]);
+
+            const blockSize = config.getRawConfig().gerbil.site_block_size;
+            const subnets = [...sitesQuery, ...clientsQuery]
+                .map((row) => row.subnet)
+                .filter(
+                    (subnet): subnet is string =>
+                        !!subnet &&
+                        /^(\d{1,3}\.){3}\d{1,3}\/\d{1,2}$/.test(subnet)
+                );
+            subnets.push(exitNode.address.replace(/\/\d+$/, `/${blockSize}`));
+
+            return findNextAvailableCidr(subnets, blockSize, exitNode.address);
+        },
+        5000 // 5 second lock TTL - subnet allocation should be quick
+    );
+}
@@ -2,3 +2,5 @@ export * from "./exitNodes";
 export * from "./exitNodeComms";
 export * from "./subnet";
 export * from "./getCurrentExitNodeId";
+export * from "./calculateExitNodeWeight";
+export * from "./getUniqueSubnetForExitNode";
@@ -528,7 +528,10 @@ export function generateRemoteSubnets(

 export type Alias = { alias: string | null; aliasAddress: string | null };

-export function generateAliasConfig(allSiteResources: SiteResource[]): Alias[] {
+export function generateAliasConfig(
+    allSiteResources: SiteResource[],
+    overrideIp?: string
+): Alias[] {
    return allSiteResources
        .filter(
            (sr) =>
@@ -539,7 +542,7 @@ export function generateAliasConfig(allSiteResources: SiteResource[]): Alias[] {
        )
        .map((sr) => ({
            alias: sr.alias || sr.fullDomain,
-            aliasAddress: sr.aliasAddress
+            aliasAddress: overrideIp || sr.aliasAddress
        }));
 }

@@ -660,9 +663,10 @@ export type CertRef = { id: string; cert: string; key: string };
 * certificate (e.g. a wildcard cert used by thousands of site resources)
 * only need that certificate sent once per sync message.
 */
-export function dedupeCertsForTargets(
-    targetsV2: SubnetProxyTargetV2[]
-): { targets: SubnetProxyTargetV2[]; certs: CertRef[] } {
+export function dedupeCertsForTargets(targetsV2: SubnetProxyTargetV2[]): {
+    targets: SubnetProxyTargetV2[];
+    certs: CertRef[];
+} {
    const idByContent = new Map<string, string>();
    const certs: CertRef[] = [];

@@ -674,7 +678,10 @@ export function dedupeCertsForTargets(
        const contentKey = `${target.tlsCert}|${target.tlsKey}`;
        let id = idByContent.get(contentKey);
        if (!id) {
-            id = createHash("sha1").update(contentKey).digest("hex").slice(0, 16);
+            id = createHash("sha1")
+                .update(contentKey)
+                .digest("hex")
+                .slice(0, 16);
            idByContent.set(contentKey, id);
            certs.push({ id, cert: target.tlsCert, key: target.tlsKey });
        }
@@ -708,7 +715,9 @@ export async function batchFetchCertsForSiteResources(
 ): Promise<CertByDomain> {
    const domains = new Set(
        allSiteResources
-            .filter((r) => r.enabled && r.mode === "http" && r.ssl && r.fullDomain)
+            .filter(
+                (r) => r.enabled && r.mode === "http" && r.ssl && r.fullDomain
+            )
            .map((r) => r.fullDomain as string)
    );

@@ -852,7 +861,11 @@ export async function generateSubnetProxyTargetV2(
                        new Set([siteResource.fullDomain]),
                        true
                    );
-                    if (certs.length > 0 && certs[0].certFile && certs[0].keyFile) {
+                    if (
+                        certs.length > 0 &&
+                        certs[0].certFile &&
+                        certs[0].keyFile
+                    ) {
                        tlsCert = certs[0].certFile;
                        tlsKey = certs[0].keyFile;
                    } else {
@@ -79,7 +79,13 @@ export const configSchema = z
                    .default(3001)
                    .transform(stoi)
                    .pipe(portSchema),
+                ai_gateway_port: portSchema
+                    .optional()
+                    .default(3005)
+                    .transform(stoi)
+                    .pipe(portSchema),
                badger_override: z.string().optional(),
+                ai_gateway_override: z.string().optional(),
                next_port: portSchema
                    .optional()
                    .default(3002)
@@ -139,6 +145,7 @@ export const configSchema = z
                integration_port: 3003,
                external_port: 3000,
                internal_port: 3001,
+                ai_gateway_port: 3005,
                next_port: 3002,
                internal_hostname: "pangolin",
                session_cookie_name: "p_session_token",
@@ -19,7 +19,7 @@ import {
    userOrgRoles,
    userSiteResources
 } from "@server/db";
-import { and, count, eq, inArray, ne } from "drizzle-orm";
+import { and, count, eq, inArray, isNotNull, ne } from "drizzle-orm";

 import { deletePeersBatch as newtDeletePeersBatch } from "@server/routers/newt/peers";
 import {
@@ -27,6 +27,9 @@ import {
    deletePeersBatch as olmDeletePeersBatch
 } from "@server/routers/olm/peers";
 import { sendToExitNode } from "#dynamic/lib/exitNodes";
+import { sendToClientsBatch } from "#dynamic/routers/ws";
+import { canCompress } from "@server/lib/clientVersionChecks";
+import config from "@server/lib/config";
 import logger from "@server/logger";
 import {
    generateAliasConfig,
@@ -187,7 +190,12 @@ export async function getClientSiteResourceAccess(
        `rebuildClientAssociations: [getClientSiteResourceAccess] siteResourceId=${siteResource.siteResourceId} networkId=${siteResource.networkId} siteCount=${sitesList.length} siteIds=[${sitesList.map((s) => s.siteId).join(", ")}]`
    );

-    if (sitesList.length === 0) {
+    if (sitesList.length === 0 && siteResource.networkId !== null) {
+        // A site resource with a networkId is expected to have at least one
+        // site attached via siteNetworks. Resources with no networkId (e.g.
+        // inference-mode resources, which connect clients directly to the
+        // exit node instead of any site) are expected to have no sites, so
+        // don't warn for those.
        logger.warn(
            `No sites found for siteResource ${siteResource.siteResourceId} with networkId ${siteResource.networkId}`
        );
@@ -687,6 +695,22 @@ async function rebuildClientAssociationsFromSiteResourceImpl(
        clientSiteResourcesToRemove,
        trx
    );
+
+    // If this resource requires clients to be connected to the exit node
+    // (e.g. an inference resource), re-sync the connect/disconnect state for
+    // every client whose access to it may have changed - both those who
+    // currently have access and those who just lost it.
+    if (siteResource.requiresExitNodeConnection) {
+        await syncClientExitNodeConnections(
+            Array.from(
+                new Set([
+                    ...mergedAllClientIds,
+                    ...existingClientSiteResourceIds
+                ])
+            ),
+            trx
+        );
+    }
 }

 async function handleMessagesForSiteClients(
@@ -966,7 +990,7 @@ export async function updateClientSiteDestinations(
        .where(eq(clientSitesAssociationsCache.clientId, client.clientId));

    for (const site of sitesData) {
-        if (!site.sites.subnet) {
+        if (!site.sites.exitNodeSubnet) {
            logger.debug(`Site ${site.sites.siteId} has no subnet, skipping`);
            continue;
        }
@@ -1002,7 +1026,7 @@ export async function updateClientSiteDestinations(
                sourcePort: parsedEndpoint.port,
                destinations: [
                    {
-                        destinationIP: site.sites.subnet.split("/")[0],
+                        destinationIP: site.sites.exitNodeSubnet.split("/")[0],
                        destinationPort: site.sites.listenPort || 1 // this satisfies gerbil for now but should be reevaluated
                    }
                ]
@@ -1010,7 +1034,7 @@ export async function updateClientSiteDestinations(
        } else {
            // add to the existing destinations
            destinations.destinations.push({
-                destinationIP: site.sites.subnet.split("/")[0],
+                destinationIP: site.sites.exitNodeSubnet.split("/")[0],
                destinationPort: site.sites.listenPort || 1 // this satisfies gerbil for now but should be reevaluated
            });
        }
@@ -1052,6 +1076,265 @@ export async function updateClientSiteDestinations(
    }
 }

+// Determines, for each of the given clients, whether they currently have
+// access to any enabled site resource with requiresExitNodeConnection set
+// (e.g. an inference-mode resource) and tells the client's olm to connect to
+// or disconnect from its assigned exit node accordingly. Site resources with
+// requiresExitNodeConnection don't belong to any site/network, so this can't
+// be derived from the per-site peer logic above - it has to be recomputed
+// from the client's full current resource access every time that access
+// changes.
+async function syncClientExitNodeConnections(
+    clientIds: number[],
+    trx: Transaction | typeof db = db
+): Promise<void> {
+    const uniqueClientIds = Array.from(new Set(clientIds));
+    if (uniqueClientIds.length === 0) {
+        return;
+    }
+
+    // Only clients with an exit node assigned can be told to connect/disconnect.
+    const clientsData = await trx
+        .select({
+            clientId: clients.clientId,
+            exitNodeId: clients.exitNodeId,
+            exitNodeSubnet: clients.exitNodeSubnet
+        })
+        .from(clients)
+        .where(
+            and(
+                inArray(clients.clientId, uniqueClientIds),
+                isNotNull(clients.exitNodeId)
+            )
+        );
+
+    if (clientsData.length === 0) {
+        return;
+    }
+
+    const clientIdsWithExitNode = clientsData.map((c) => c.clientId);
+
+    const requiresExitNodeRows = await trx
+        .select({
+            clientId: clientSiteResourcesAssociationsCache.clientId,
+            alias: siteResources.alias,
+            fullDomain: siteResources.fullDomain
+        })
+        .from(clientSiteResourcesAssociationsCache)
+        .innerJoin(
+            siteResources,
+            eq(
+                clientSiteResourcesAssociationsCache.siteResourceId,
+                siteResources.siteResourceId
+            )
+        )
+        .where(
+            and(
+                inArray(
+                    clientSiteResourcesAssociationsCache.clientId,
+                    clientIdsWithExitNode
+                ),
+                eq(siteResources.enabled, true),
+                eq(siteResources.requiresExitNodeConnection, true)
+            )
+        );
+
+    const needsConnectSet = new Set(
+        requiresExitNodeRows.map((r) => r.clientId)
+    );
+
+    // Aliases for every exit-node-backed resource this client can reach, so
+    // the live connect push carries the same alias list the register/reconnect
+    // path (buildSiteConfigurationForOlmClient) would compute.
+    const exitNodeAliasesByClientId = new Map<number, (string | null)[]>();
+    for (const row of requiresExitNodeRows) {
+        if (row.alias == null && row.fullDomain == null) continue;
+        const existing = exitNodeAliasesByClientId.get(row.clientId);
+        if (existing) {
+            existing.push(row.fullDomain || row.alias); // accept both for now in case we have other resource types that dont use the full domain
+        } else {
+            exitNodeAliasesByClientId.set(row.clientId, [
+                row.fullDomain || row.alias
+            ]);
+        }
+    }
+
+    const exitNodeIds = Array.from(
+        new Set(
+            clientsData
+                .map((c) => c.exitNodeId)
+                .filter((id): id is number => id !== null)
+        )
+    );
+
+    const exitNodeRows =
+        exitNodeIds.length > 0
+            ? await trx
+                  .select()
+                  .from(exitNodes)
+                  .where(inArray(exitNodes.exitNodeId, exitNodeIds))
+            : [];
+    const exitNodeById = new Map(exitNodeRows.map((n) => [n.exitNodeId, n]));
+
+    const olmRows = await trx
+        .select({
+            clientId: olms.clientId,
+            olmId: olms.olmId,
+            version: olms.version
+        })
+        .from(olms)
+        .where(inArray(olms.clientId, clientIdsWithExitNode));
+    const olmByClientId = new Map(
+        olmRows
+            .filter((r) => r.clientId !== null)
+            .map((r) => [r.clientId as number, r])
+    );
+
+    const relayPort = config.getRawConfig().gerbil.clients_start_port;
+
+    const connectPayloads: {
+        clientId: string;
+        message: { type: string; data: any };
+        options: { compress: boolean; incrementConfigVersion: boolean };
+    }[] = [];
+    const disconnectPayloads: {
+        clientId: string;
+        message: { type: string; data: any };
+        options: { compress: boolean; incrementConfigVersion: boolean };
+    }[] = [];
+
+    for (const client of clientsData) {
+        const olm = olmByClientId.get(client.clientId);
+        if (!olm) {
+            // No olm registered for this client yet/anymore, nothing to send.
+            continue;
+        }
+
+        const needsConnect = needsConnectSet.has(client.clientId);
+
+        if (needsConnect) {
+            const exitNode = client.exitNodeId
+                ? exitNodeById.get(client.exitNodeId)
+                : undefined;
+            if (!exitNode || !client.exitNodeSubnet) {
+                logger.warn(
+                    `rebuildClientAssociations: [syncClientExitNodeConnections] client ${client.clientId} needs an exit node connection but has no exit node or subnet assigned`
+                );
+                continue;
+            }
+
+            connectPayloads.push({
+                clientId: olm.olmId,
+                message: {
+                    type: "olm/wg/exitnode/connect",
+                    data: {
+                        connect: true,
+                        endpoint: `${exitNode.endpoint}:${exitNode.listenPort}`,
+                        relayPort,
+                        publicKey: exitNode.publicKey,
+                        serverIP: exitNode.address.split("/")[0],
+                        tunnelIP: client.exitNodeSubnet.split("/")[0],
+                        aliases:
+                            exitNodeAliasesByClientId.get(client.clientId) ?? []
+                    }
+                },
+                options: {
+                    compress: canCompress(olm.version, "olm"),
+                    incrementConfigVersion: true
+                }
+            });
+        } else {
+            disconnectPayloads.push({
+                clientId: olm.olmId,
+                message: {
+                    type: "olm/wg/exitnode/disconnect",
+                    data: {}
+                },
+                options: {
+                    compress: canCompress(olm.version, "olm"),
+                    incrementConfigVersion: true
+                }
+            });
+        }
+    }
+
+    if (connectPayloads.length > 0) {
+        await sendToClientsBatch(connectPayloads).catch((error) => {
+            logger.error(
+                `rebuildClientAssociations: Error sending exit node connect messages:`,
+                error
+            );
+        });
+    }
+
+    if (disconnectPayloads.length > 0) {
+        await sendToClientsBatch(disconnectPayloads).catch((error) => {
+            logger.error(
+                `rebuildClientAssociations: Error sending exit node disconnect messages:`,
+                error
+            );
+        });
+    }
+}
+
+// Notifies the olms of every given client that the alias of the site resource
+// they're using an exit node connection for has changed, via the dedicated
+// exit node data-update message. Unlike syncClientExitNodeConnections, this
+// doesn't touch connect/disconnect state - it's purely a rename for clients
+// that are (and remain) connected to the exit node for this resource.
+async function syncClientExitNodeAliasUpdate(
+    clientIds: number[],
+    oldAlias: string | null,
+    newAlias: string | null,
+    trx: Transaction | typeof db = db
+): Promise<void> {
+    const uniqueClientIds = Array.from(new Set(clientIds));
+    if (uniqueClientIds.length === 0) {
+        return;
+    }
+
+    const oldAliases = oldAlias ? [oldAlias] : [];
+    const newAliases = newAlias ? [newAlias] : [];
+    if (oldAliases.length === 0 && newAliases.length === 0) {
+        return;
+    }
+
+    const olmRows = await trx
+        .select({
+            clientId: olms.clientId,
+            olmId: olms.olmId,
+            version: olms.version
+        })
+        .from(olms)
+        .where(inArray(olms.clientId, uniqueClientIds));
+
+    const updatePayloads = olmRows
+        .filter((r) => r.clientId !== null)
+        .map((olm) => ({
+            clientId: olm.olmId,
+            message: {
+                type: "olm/wg/exitnode/data/update",
+                data: {
+                    oldAliases,
+                    newAliases
+                }
+            },
+            options: {
+                compress: canCompress(olm.version, "olm"),
+                incrementConfigVersion: true // this is important information we would need to sync
+            }
+        }));
+
+    if (updatePayloads.length > 0) {
+        await sendToClientsBatch(updatePayloads).catch((error) => {
+            logger.error(
+                `rebuildClientAssociations: Error sending exit node alias update messages:`,
+                error
+            );
+        });
+    }
+}
+
 async function handleSubnetProxyTargetUpdates(
    siteResource: SiteResource,
    sitesList: Site[],
@@ -1282,7 +1565,7 @@ export async function handleMessagingForUpdatedSiteResource(
        `handleMessagingForUpdatedSiteResource: fetched newts for ${newtsForSites.length}/${allSiteIds.length} site(s)`
    );

-    // WARNING: THIS RELIES ON THE CACHE TABLES BEING UP TO DATE, SO CALL THIS AFTER THE ASSOCIATION CACHE IS UPDATED
+    // !!!!!!!!!!!!!!!!!! WARNING: THIS RELIES ON THE CACHE TABLES BEING UP TO DATE, SO CALL THIS AFTER THE ASSOCIATION CACHE IS UPDATED !!!!!!!!!!!!!!!!!!
    const mergedAllClients = await trx
        .select({
            clientId: clientSiteResourcesAssociationsCache.clientId,
@@ -1709,6 +1992,38 @@ export async function handleMessagingForUpdatedSiteResource(
        );
    }

+    // For a resource that stays on an exit node connection across the update,
+    // the alias is the only field that affects already-connected clients (the
+    // exit node itself, its endpoint, etc. are not per-resource). Tell those
+    // clients' olms about the rename directly via the exit node data-update
+    // message rather than a full connect/disconnect cycle.
+    if (
+        existingSiteResource?.requiresExitNodeConnection &&
+        updatedSiteResource.requiresExitNodeConnection &&
+        aliasChanged
+    ) {
+        await syncClientExitNodeAliasUpdate(
+            mergedAllClients.map((c) => c.clientId),
+            existingSiteResource.alias,
+            updatedSiteResource.alias,
+            trx
+        );
+    }
+
+    // If this resource requires (or required) clients to be connected to the
+    // exit node (e.g. an inference resource), re-sync connect/disconnect
+    // state for every client currently associated with it - covers toggling
+    // requiresExitNodeConnection on update as well as enabling/disabling it.
+    if (
+        updatedSiteResource.requiresExitNodeConnection ||
+        existingSiteResource?.requiresExitNodeConnection
+    ) {
+        await syncClientExitNodeConnections(
+            mergedAllClients.map((c) => c.clientId),
+            trx
+        );
+    }
+
    logger.debug(
        `handleMessagingForUpdatedSiteResource: DONE siteResourceId=${updatedSiteResource.siteResourceId}`
    );
@@ -1990,6 +2305,10 @@ async function rebuildClientAssociationsFromClientImpl(
        resourcesToRemove,
        trx
    );
+
+    // Re-sync exit node connect/disconnect state based on this client's
+    // current full set of resource access (e.g. inference resources).
+    await syncClientExitNodeConnections([client.clientId], trx);
 }

 async function handleMessagesForClientSites(
@@ -516,6 +516,11 @@ export class TraefikConfigManager {
            const maintenanceHost =
                config.getRawConfig().server.internal_hostname;
            const pangolinUIUrl = `http://${maintenanceHost}:${maintenancePort}`;
+            const aiGatewayUrl =
+                config.getRawConfig().server.ai_gateway_override ||
+                `http://${maintenanceHost}:${
+                    config.getRawConfig().server.ai_gateway_port
+                }`;

            // logger.debug(`Fetching traefik config for exit node: ${currentExitNode}`);
            traefikConfig = await getTraefikConfig(
@@ -528,7 +533,8 @@ export class TraefikConfigManager {
                    ? false
                    : config.getRawConfig().traefik.allow_raw_resources, // dont allow raw resources on saas otherwise use config
                pangolinUIUrl, // generate maintenance pages on cloud and hybrid
-                pangolinUIUrl // generate browser gateway targets on cloud and hybrid
+                pangolinUIUrl, // generate browser gateway targets on cloud and hybrid
+                aiGatewayUrl
            );

            const domains = new Set<string>();
@@ -1,4 +1,12 @@
-import { db, targetHealthCheck, domains } from "@server/db";
+import {
+    db,
+    targetHealthCheck,
+    domains,
+    aiProviders,
+    resourceAiProviders,
+    siteResources,
+    exitNodes
+} from "@server/db";
 import {
    and,
    eq,
@@ -15,6 +23,7 @@ import config from "@server/lib/config";
 import { resources, sites, Target, targets } from "@server/db";
 import createPathRewriteMiddleware from "./middleware";
 import { sanitize, encodePath, validatePathRewriteConfig } from "./utils";
+import regionalCache from "@server/lib/cache";

 const redirectHttpsMiddlewareName = "redirect-to-https";
 const badgerMiddlewareName = "badger";
@@ -45,8 +54,24 @@ export async function getTraefikConfig(
    generateLoginPageRouters = false, // UNUSED BUT USED IN PRIVATE
    allowRawResources = true,
    maintenancePageUiUrl: string | null = null, // UNUSED BUT USED IN PRIVATE
-    browserGatewayUiUrl: string | null = null // UNUSED BUT USED IN PRIVATE
+    browserGatewayUiUrl: string | null = null, // UNUSED BUT USED IN PRIVATE
+    aiGatewayUrl: string | null = null
 ): Promise<any> {
+    // Get the exit node but cache it for 5 minutes to avoid hitting the DB too often
+    const exitNodeCacheKey = `exitNode:${exitNodeId}`;
+    let exitNode =
+        await regionalCache.get<typeof exitNodes.$inferSelect>(
+            exitNodeCacheKey
+        );
+    if (!exitNode) {
+        [exitNode] = await db
+            .select()
+            .from(exitNodes)
+            .where(eq(exitNodes.exitNodeId, exitNodeId))
+            .limit(1);
+        await regionalCache.set(exitNodeCacheKey, exitNode, 300);
+    }
+
    // Get resources with their targets and sites in a single optimized query
    // Start from sites on this exit node, then join to targets and resources
    const resourcesWithTargetsAndSites = await db
@@ -87,7 +112,7 @@ export async function getTraefikConfig(
            siteId: sites.siteId,
            siteType: sites.type,
            siteOnline: sites.online,
-            subnet: sites.subnet,
+            subnet: sites.exitNodeSubnet,
            exitNodeId: sites.exitNodeId,
            // Domain cert resolver fields
            domainCertResolver: domains.certResolver,
@@ -209,8 +234,42 @@ export async function getTraefikConfig(
        });
    });

+    // Inference-mode resources have no targets/sites (their "backend" is the
+    // central AI gateway), so they can't be reached via the targets->sites
+    // join above - query them separately and include them on every exit node.
+    const inferenceResources = await db
+        .selectDistinct({
+            resourceId: resources.resourceId,
+            resourceName: resources.name,
+            fullDomain: resources.fullDomain,
+            ssl: resources.ssl,
+            subdomain: resources.subdomain,
+            domainId: resources.domainId,
+            enabled: resources.enabled,
+            wildcard: resources.wildcard,
+            domainCertResolver: domains.certResolver,
+            preferWildcardCert: domains.preferWildcardCert
+        })
+        .from(resources)
+        // .innerJoin(
+        //     resourceAiProviders,
+        //     eq(resources.resourceId, resourceAiProviders.resourceId)
+        // )
+        // .innerJoin(
+        //     aiProviders,
+        //     eq(resourceAiProviders.providerId, aiProviders.providerId)
+        // )
+        .leftJoin(domains, eq(domains.domainId, resources.domainId))
+        .where(
+            and(
+                eq(resources.mode, "inference"),
+                eq(resources.enabled, true)
+                // eq(aiProviders.enabled, true)
+            )
+        );
+
    // make sure we have at least one resource
-    if (resourcesMap.size === 0) {
+    if (resourcesMap.size === 0 && inferenceResources.length === 0) {
        return {};
    }

@@ -673,5 +732,229 @@ export async function getTraefikConfig(
            };
        }
    }
+
+    if (aiGatewayUrl) {
+        // The AI gateway may live on a different host than the inference
+        // resource itself (e.g. a remote exit node forwarding to the
+        // central dashboard over a tunnel). passHostHeader would forward
+        // the resource's own Host, which that external host won't
+        // recognize, so we pin the Host header to the gateway's own host
+        // and smuggle the original resource host through in "p-host"
+        // instead.
+        let aiGatewayHost: string | undefined;
+        try {
+            aiGatewayHost = new URL(aiGatewayUrl).host;
+        } catch {
+            aiGatewayHost = undefined;
+        }
+
+        // Public inference resources: same TLS/cert-resolver handling as
+        // plain http-mode resources, but the service points at the AI
+        // gateway instead of any real backend targets.
+        for (const ir of inferenceResources) {
+            if (!ir.enabled) continue;
+            if (!ir.domainId || !ir.fullDomain) continue;
+
+            if (!config_output.http.routers) config_output.http.routers = {};
+            if (!config_output.http.services) config_output.http.services = {};
+
+            const fullDomain = ir.fullDomain;
+            const irKey = `inference-r${ir.resourceId}`;
+            const routerName = `${irKey}-router`;
+            const serviceName = `${irKey}-service`;
+
+            let rule: string;
+            if (ir.wildcard && fullDomain.startsWith("*.")) {
+                const escaped = fullDomain.slice(2).replace(/\./g, "\\.");
+                rule = `HostRegexp(\`^[^.]+\\.${escaped}$\`)`;
+            } else {
+                rule = `Host(\`${fullDomain}\`)`;
+            }
+
+            const domainParts = fullDomain.split(".");
+            let wildCard;
+            if (domainParts.length <= 2) {
+                wildCard = `*.${domainParts.join(".")}`;
+            } else {
+                wildCard = `*.${domainParts.slice(1).join(".")}`;
+            }
+            if (!ir.subdomain) {
+                wildCard = fullDomain;
+            }
+
+            const globalDefaultResolver =
+                config.getRawConfig().traefik.cert_resolver;
+            const globalDefaultPreferWildcard =
+                config.getRawConfig().traefik.prefer_wildcard_cert;
+            const resolverName = ir.domainCertResolver
+                ? ir.domainCertResolver.trim()
+                : globalDefaultResolver;
+            const preferWildcard =
+                ir.preferWildcardCert !== undefined &&
+                ir.preferWildcardCert !== null
+                    ? ir.preferWildcardCert
+                    : globalDefaultPreferWildcard;
+
+            const tls = {
+                certResolver: resolverName,
+                ...(preferWildcard ? { domains: [{ main: wildCard }] } : {})
+            };
+
+            const irHeadersMiddlewareName = `${irKey}-headers-middleware`;
+            if (!config_output.http.middlewares) {
+                config_output.http.middlewares = {};
+            }
+            config_output.http.middlewares[irHeadersMiddlewareName] = {
+                headers: {
+                    customRequestHeaders: {
+                        ...(aiGatewayHost ? { Host: aiGatewayHost } : {}),
+                        "p-host": fullDomain
+                    }
+                }
+            };
+
+            const additionalMiddlewares =
+                config.getRawConfig().traefik.additional_middlewares || [];
+            const routerMiddlewares = [
+                badgerMiddlewareName,
+                irHeadersMiddlewareName,
+                ...additionalMiddlewares
+            ];
+
+            if (ir.ssl) {
+                config_output.http.routers[routerName + "-redirect"] = {
+                    entryPoints: [
+                        config.getRawConfig().traefik.http_entrypoint
+                    ],
+                    middlewares: [redirectHttpsMiddlewareName],
+                    service: serviceName,
+                    rule,
+                    priority: 100
+                };
+            }
+
+            config_output.http.routers[routerName] = {
+                entryPoints: [
+                    ir.ssl
+                        ? config.getRawConfig().traefik.https_entrypoint
+                        : config.getRawConfig().traefik.http_entrypoint
+                ],
+                middlewares: routerMiddlewares,
+                service: serviceName,
+                rule,
+                priority: 100,
+                ...(ir.ssl ? { tls } : {})
+            };
+
+            config_output.http.services[serviceName] = {
+                loadBalancer: {
+                    servers: [{ url: aiGatewayUrl }]
+                }
+            };
+        }
+
+        // Private (siteResource) inference resources: routed by their alias
+        // instead of a public fullDomain, and deliberately WITHOUT the
+        // badger middleware - no per-user auth/policy stack exists for
+        // siteResources today, so gating here is reachability-only for now.
+        const siteResourcesInference = await db
+            .selectDistinct({
+                siteResourceId: siteResources.siteResourceId,
+                fullDomain: siteResources.fullDomain,
+                ssl: siteResources.ssl,
+                enabled: siteResources.enabled
+            })
+            .from(siteResources)
+            .where(
+                and(
+                    eq(siteResources.mode, "inference"),
+                    eq(siteResources.enabled, true),
+                    isNotNull(siteResources.fullDomain)
+                )
+            );
+
+        for (const sr of siteResourcesInference) {
+            if (!sr.enabled || !sr.fullDomain) continue;
+
+            if (!config_output.http.routers) config_output.http.routers = {};
+            if (!config_output.http.services) config_output.http.services = {};
+
+            const fullDomain = sr.fullDomain;
+            const srKey = `inference-sr${sr.siteResourceId}`;
+            const routerName = `${srKey}-router`;
+            const serviceName = `${srKey}-service`;
+            const rule = `Host(\`${fullDomain}\`) && ClientIP(${exitNode.address})`; // restrict to coming from the exit node ip range that the client is connected to
+
+            const domainParts = fullDomain.split(".");
+            const wildCard =
+                domainParts.length <= 2
+                    ? `*.${domainParts.join(".")}`
+                    : `*.${domainParts.slice(1).join(".")}`;
+
+            const globalDefaultResolver =
+                config.getRawConfig().traefik.cert_resolver;
+            const globalDefaultPreferWildcard =
+                config.getRawConfig().traefik.prefer_wildcard_cert;
+
+            const tls = {
+                certResolver: globalDefaultResolver,
+                ...(globalDefaultPreferWildcard
+                    ? { domains: [{ main: wildCard }] }
+                    : {})
+            };
+
+            const srHeadersMiddlewareName = `${srKey}-headers-middleware`;
+            if (!config_output.http.middlewares) {
+                config_output.http.middlewares = {};
+            }
+            config_output.http.middlewares[srHeadersMiddlewareName] = {
+                headers: {
+                    customRequestHeaders: {
+                        ...(aiGatewayHost ? { Host: aiGatewayHost } : {}),
+                        "p-host": fullDomain
+                    }
+                }
+            };
+
+            const additionalMiddlewares =
+                config.getRawConfig().traefik.additional_middlewares || [];
+            const routerMiddlewares = [
+                srHeadersMiddlewareName,
+                ...additionalMiddlewares
+            ];
+
+            if (sr.ssl) {
+                config_output.http.routers[routerName + "-redirect"] = {
+                    entryPoints: [
+                        config.getRawConfig().traefik.http_entrypoint
+                    ],
+                    middlewares: [redirectHttpsMiddlewareName],
+                    service: serviceName,
+                    rule,
+                    priority: 100
+                };
+            }
+
+            config_output.http.routers[routerName] = {
+                entryPoints: [
+                    sr.ssl
+                        ? config.getRawConfig().traefik.https_entrypoint
+                        : config.getRawConfig().traefik.http_entrypoint
+                ],
+                middlewares: routerMiddlewares,
+                service: serviceName,
+                rule,
+                priority: 100,
+                ...(sr.ssl ? { tls } : {})
+            };
+
+            config_output.http.services[serviceName] = {
+                loadBalancer: {
+                    servers: [{ url: aiGatewayUrl }]
+                }
+            };
+        }
+    }
+
    return config_output;
 }
@@ -27,6 +27,8 @@ export * from "./verifyUserHasAction";
 export * from "./verifyApiKeyAccess";
 export * from "./verifySiteProvisioningKeyAccess";
 export * from "./verifyDomainAccess";
+export * from "./verifyAiProviderAccess";
+export * from "./verifyAiModelAccess";
 export * from "./verifyUserIsOrgOwner";
 export * from "./verifyUserFromResourceSession";
 export * from "./verifySiteResourceAccess";
@@ -16,5 +16,7 @@ export * from "./verifyApiKeyClientAccess";
 export * from "./verifyApiKeySiteResourceAccess";
 export * from "./verifyApiKeyIdpAccess";
 export * from "./verifyApiKeyDomainAccess";
+export * from "./verifyApiKeyAiProviderAccess";
+export * from "./verifyApiKeyAiModelAccess";
 export * from "./verifyApiKeyResourcePolicyAccess";
 export * from "./verifyApiKeySiteProvisioningKeyAccess";
@@ -0,0 +1,94 @@
+import { Request, Response, NextFunction } from "express";
+import { aiModels, aiProviders, apiKeyOrg, db } from "@server/db";
+import { and, eq } from "drizzle-orm";
+import createHttpError from "http-errors";
+import HttpCode from "@server/types/HttpCode";
+import { getFirstString } from "@server/lib/requestParams";
+
+export async function verifyApiKeyAiModelAccess(
+    req: Request,
+    res: Response,
+    next: NextFunction
+) {
+    try {
+        const apiKey = req.apiKey;
+        const modelIdRaw = getFirstString(req.params.modelId);
+        const modelId = Number.parseInt(modelIdRaw ?? "", 10);
+
+        if (!apiKey) {
+            return next(
+                createHttpError(HttpCode.UNAUTHORIZED, "Key not authenticated")
+            );
+        }
+
+        if (Number.isNaN(modelId)) {
+            return next(
+                createHttpError(HttpCode.BAD_REQUEST, "Invalid model ID")
+            );
+        }
+
+        const [row] = await db
+            .select({
+                model: aiModels,
+                provider: aiProviders
+            })
+            .from(aiModels)
+            .innerJoin(
+                aiProviders,
+                eq(aiModels.providerId, aiProviders.providerId)
+            )
+            .where(eq(aiModels.modelId, modelId))
+            .limit(1);
+
+        if (!row) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    `AI model with ID ${modelId} not found`
+                )
+            );
+        }
+
+        if (apiKey.isRoot) {
+            req.aiProvider = row.provider;
+            req.aiModel = row.model;
+            return next();
+        }
+
+        const orgId = row.provider.orgId;
+
+        if (!req.apiKeyOrg || req.apiKeyOrg.orgId !== orgId) {
+            const apiKeyOrgRes = await db
+                .select()
+                .from(apiKeyOrg)
+                .where(
+                    and(
+                        eq(apiKeyOrg.apiKeyId, apiKey.apiKeyId),
+                        eq(apiKeyOrg.orgId, orgId)
+                    )
+                )
+                .limit(1);
+            req.apiKeyOrg = apiKeyOrgRes[0];
+        }
+
+        if (!req.apiKeyOrg) {
+            return next(
+                createHttpError(
+                    HttpCode.FORBIDDEN,
+                    "Key does not have access to this organization"
+                )
+            );
+        }
+
+        req.aiProvider = row.provider;
+        req.aiModel = row.model;
+        return next();
+    } catch (error) {
+        return next(
+            createHttpError(
+                HttpCode.INTERNAL_SERVER_ERROR,
+                "Error verifying AI model access"
+            )
+        );
+    }
+}
@@ -0,0 +1,85 @@
+import { Request, Response, NextFunction } from "express";
+import { aiProviders, apiKeyOrg, db } from "@server/db";
+import { and, eq } from "drizzle-orm";
+import createHttpError from "http-errors";
+import HttpCode from "@server/types/HttpCode";
+import { getFirstString } from "@server/lib/requestParams";
+
+export async function verifyApiKeyAiProviderAccess(
+    req: Request,
+    res: Response,
+    next: NextFunction
+) {
+    try {
+        const apiKey = req.apiKey;
+        const providerIdRaw = getFirstString(req.params.providerId);
+        const providerId = Number.parseInt(providerIdRaw ?? "", 10);
+
+        if (!apiKey) {
+            return next(
+                createHttpError(HttpCode.UNAUTHORIZED, "Key not authenticated")
+            );
+        }
+
+        if (Number.isNaN(providerId)) {
+            return next(
+                createHttpError(HttpCode.BAD_REQUEST, "Invalid provider ID")
+            );
+        }
+
+        const [provider] = await db
+            .select()
+            .from(aiProviders)
+            .where(eq(aiProviders.providerId, providerId))
+            .limit(1);
+
+        if (!provider) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    `AI provider with ID ${providerId} not found`
+                )
+            );
+        }
+
+        if (apiKey.isRoot) {
+            req.aiProvider = provider;
+            return next();
+        }
+
+        const orgId = provider.orgId;
+
+        if (!req.apiKeyOrg || req.apiKeyOrg.orgId !== orgId) {
+            const apiKeyOrgRes = await db
+                .select()
+                .from(apiKeyOrg)
+                .where(
+                    and(
+                        eq(apiKeyOrg.apiKeyId, apiKey.apiKeyId),
+                        eq(apiKeyOrg.orgId, orgId)
+                    )
+                )
+                .limit(1);
+            req.apiKeyOrg = apiKeyOrgRes[0];
+        }
+
+        if (!req.apiKeyOrg) {
+            return next(
+                createHttpError(
+                    HttpCode.FORBIDDEN,
+                    "Key does not have access to this organization"
+                )
+            );
+        }
+
+        req.aiProvider = provider;
+        return next();
+    } catch (error) {
+        return next(
+            createHttpError(
+                HttpCode.INTERNAL_SERVER_ERROR,
+                "Error verifying AI provider access"
+            )
+        );
+    }
+}
@@ -1,6 +1,6 @@
 import { Request, Response, NextFunction } from "express";
 import { db } from "@server/db";
-import { resources, targets, apiKeyOrg } from "@server/db";
+import { aiProviders, resources, targets, apiKeyOrg } from "@server/db";
 import { and, eq } from "drizzle-orm";
 import createHttpError from "http-errors";
 import HttpCode from "@server/types/HttpCode";
@@ -43,43 +43,65 @@ export async function verifyApiKeyTargetAccess(
            );
        }

-        const resourceId = target.resourceId;
-        if (!resourceId) {
+        const { resourceId, providerId } = target;
+        if ((!resourceId && !providerId) || (resourceId && providerId)) {
            return next(
                createHttpError(
                    HttpCode.INTERNAL_SERVER_ERROR,
-                    `Target with ID ${targetId} does not have a resource ID`
-                )
-            );
-        }
-
-        const [resource] = await db
-            .select()
-            .from(resources)
-            .where(eq(resources.resourceId, resourceId))
-            .limit(1);
-
-        if (!resource) {
-            return next(
-                createHttpError(
-                    HttpCode.NOT_FOUND,
-                    `Resource with ID ${resourceId} not found`
+                    `Target with ID ${targetId} has invalid ownership`
                )
            );
        }

        if (apiKey.isRoot) {
-            // Root keys can access any key in any org
+            // Root keys can access any target
            return next();
        }

-        if (!resource.orgId) {
-            return next(
-                createHttpError(
-                    HttpCode.INTERNAL_SERVER_ERROR,
-                    `Resource with ID ${resourceId} does not have an organization ID`
-                )
-            );
+        let orgId: string;
+        if (resourceId) {
+            const [resource] = await db
+                .select()
+                .from(resources)
+                .where(eq(resources.resourceId, resourceId))
+                .limit(1);
+
+            if (!resource) {
+                return next(
+                    createHttpError(
+                        HttpCode.NOT_FOUND,
+                        `Resource with ID ${resourceId} not found`
+                    )
+                );
+            }
+
+            if (!resource.orgId) {
+                return next(
+                    createHttpError(
+                        HttpCode.INTERNAL_SERVER_ERROR,
+                        `Resource with ID ${resourceId} does not have an organization ID`
+                    )
+                );
+            }
+
+            orgId = resource.orgId;
+        } else {
+            const [provider] = await db
+                .select()
+                .from(aiProviders)
+                .where(eq(aiProviders.providerId, providerId!))
+                .limit(1);
+
+            if (!provider) {
+                return next(
+                    createHttpError(
+                        HttpCode.NOT_FOUND,
+                        `AI provider with ID ${providerId} not found`
+                    )
+                );
+            }
+
+            orgId = provider.orgId;
        }

        if (!req.apiKeyOrg) {
@@ -89,7 +111,7 @@ export async function verifyApiKeyTargetAccess(
                .where(
                    and(
                        eq(apiKeyOrg.apiKeyId, apiKey.apiKeyId),
-                        eq(apiKeyOrg.orgId, resource.orgId)
+                        eq(apiKeyOrg.orgId, orgId)
                    )
                )
                .limit(1);
@@ -98,7 +120,7 @@ export async function verifyApiKeyTargetAccess(
            }
        }

-        if (!req.apiKeyOrg) {
+        if (!req.apiKeyOrg || req.apiKeyOrg.orgId !== orgId) {
            return next(
                createHttpError(
                    HttpCode.FORBIDDEN,
@@ -0,0 +1,107 @@
+import { Request, Response, NextFunction } from "express";
+import { aiModels, aiProviders, db, userOrgs } from "@server/db";
+import { and, eq } from "drizzle-orm";
+import createHttpError from "http-errors";
+import HttpCode from "@server/types/HttpCode";
+import { checkOrgAccessPolicy } from "#dynamic/lib/checkOrgAccessPolicy";
+import { getUserOrgRoleIds } from "@server/lib/userOrgRoles";
+import { getFirstString } from "@server/lib/requestParams";
+
+export async function verifyAiModelAccess(
+    req: Request,
+    res: Response,
+    next: NextFunction
+) {
+    try {
+        const userId = req.user!.userId;
+        const modelIdRaw = getFirstString(req.params.modelId);
+        const modelId = Number.parseInt(modelIdRaw ?? "", 10);
+
+        if (!userId) {
+            return next(
+                createHttpError(HttpCode.UNAUTHORIZED, "User not authenticated")
+            );
+        }
+
+        if (Number.isNaN(modelId)) {
+            return next(
+                createHttpError(HttpCode.BAD_REQUEST, "Invalid model ID")
+            );
+        }
+
+        const [row] = await db
+            .select({
+                model: aiModels,
+                provider: aiProviders
+            })
+            .from(aiModels)
+            .innerJoin(
+                aiProviders,
+                eq(aiModels.providerId, aiProviders.providerId)
+            )
+            .where(eq(aiModels.modelId, modelId))
+            .limit(1);
+
+        if (!row) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    `AI model with ID ${modelId} not found`
+                )
+            );
+        }
+
+        const orgId = row.provider.orgId;
+
+        if (!req.userOrg || req.userOrg.orgId !== orgId) {
+            const userOrgRole = await db
+                .select()
+                .from(userOrgs)
+                .where(
+                    and(eq(userOrgs.userId, userId), eq(userOrgs.orgId, orgId))
+                )
+                .limit(1);
+            req.userOrg = userOrgRole[0];
+        }
+
+        if (!req.userOrg) {
+            return next(
+                createHttpError(
+                    HttpCode.FORBIDDEN,
+                    "User does not have access to this organization"
+                )
+            );
+        }
+
+        if (req.orgPolicyAllowed === undefined && req.userOrg.orgId) {
+            const policyCheck = await checkOrgAccessPolicy({
+                orgId: req.userOrg.orgId,
+                userId,
+                session: req.session
+            });
+            req.orgPolicyAllowed = policyCheck.allowed;
+            if (!policyCheck.allowed || policyCheck.error) {
+                return next(
+                    createHttpError(
+                        HttpCode.FORBIDDEN,
+                        "" + (policyCheck.error || "Unknown error")
+                    )
+                );
+            }
+        }
+
+        req.userOrgId = orgId;
+        req.userOrgRoleIds = await getUserOrgRoleIds(req.userOrg.userId, orgId);
+        req.aiProvider = row.provider;
+        req.aiModel = row.model;
+
+        return next();
+    } catch (error) {
+        return next(
+            createHttpError(
+                HttpCode.INTERNAL_SERVER_ERROR,
+                "Error verifying AI model access"
+            )
+        );
+    }
+}
@@ -0,0 +1,99 @@
+import { Request, Response, NextFunction } from "express";
+import { aiProviders, db, userOrgs } from "@server/db";
+import { and, eq } from "drizzle-orm";
+import createHttpError from "http-errors";
+import HttpCode from "@server/types/HttpCode";
+import { checkOrgAccessPolicy } from "#dynamic/lib/checkOrgAccessPolicy";
+import { getUserOrgRoleIds } from "@server/lib/userOrgRoles";
+import { getFirstString } from "@server/lib/requestParams";
+
+export async function verifyAiProviderAccess(
+    req: Request,
+    res: Response,
+    next: NextFunction
+) {
+    try {
+        const userId = req.user!.userId;
+        const providerIdRaw = getFirstString(req.params.providerId);
+        const providerId = Number.parseInt(providerIdRaw ?? "", 10);
+
+        if (!userId) {
+            return next(
+                createHttpError(HttpCode.UNAUTHORIZED, "User not authenticated")
+            );
+        }
+
+        if (Number.isNaN(providerId)) {
+            return next(
+                createHttpError(HttpCode.BAD_REQUEST, "Invalid provider ID")
+            );
+        }
+
+        const [provider] = await db
+            .select()
+            .from(aiProviders)
+            .where(eq(aiProviders.providerId, providerId))
+            .limit(1);
+
+        if (!provider) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    `AI provider with ID ${providerId} not found`
+                )
+            );
+        }
+
+        const orgId = provider.orgId;
+
+        if (!req.userOrg || req.userOrg.orgId !== orgId) {
+            const userOrgRole = await db
+                .select()
+                .from(userOrgs)
+                .where(
+                    and(eq(userOrgs.userId, userId), eq(userOrgs.orgId, orgId))
+                )
+                .limit(1);
+            req.userOrg = userOrgRole[0];
+        }
+
+        if (!req.userOrg) {
+            return next(
+                createHttpError(
+                    HttpCode.FORBIDDEN,
+                    "User does not have access to this organization"
+                )
+            );
+        }
+
+        if (req.orgPolicyAllowed === undefined && req.userOrg.orgId) {
+            const policyCheck = await checkOrgAccessPolicy({
+                orgId: req.userOrg.orgId,
+                userId,
+                session: req.session
+            });
+            req.orgPolicyAllowed = policyCheck.allowed;
+            if (!policyCheck.allowed || policyCheck.error) {
+                return next(
+                    createHttpError(
+                        HttpCode.FORBIDDEN,
+                        "" + (policyCheck.error || "Unknown error")
+                    )
+                );
+            }
+        }
+
+        req.userOrgId = orgId;
+        req.userOrgRoleIds = await getUserOrgRoleIds(req.userOrg.userId, orgId);
+        req.aiProvider = provider;
+
+        return next();
+    } catch (error) {
+        return next(
+            createHttpError(
+                HttpCode.INTERNAL_SERVER_ERROR,
+                "Error verifying AI provider access"
+            )
+        );
+    }
+}
@@ -1,6 +1,6 @@
 import { Request, Response, NextFunction } from "express";
 import { db } from "@server/db";
-import { resources, targets, userOrgs } from "@server/db";
+import { aiProviders, resources, targets, userOrgs } from "@server/db";
 import { and, eq } from "drizzle-orm";
 import createHttpError from "http-errors";
 import HttpCode from "@server/types/HttpCode";
@@ -25,9 +25,7 @@ export async function verifyTargetAccess(
    }

    if (isNaN(targetId)) {
-        return next(
-            createHttpError(HttpCode.BAD_REQUEST, "Invalid organization ID")
-        );
+        return next(createHttpError(HttpCode.BAD_REQUEST, "Invalid target ID"));
    }

    const target = await db
@@ -45,73 +43,88 @@ export async function verifyTargetAccess(
        );
    }

-    const resourceId = target[0].resourceId;
+    const { resourceId, providerId } = target[0];

-    if (!resourceId) {
+    if ((!resourceId && !providerId) || (resourceId && providerId)) {
        return next(
            createHttpError(
                HttpCode.INTERNAL_SERVER_ERROR,
-                `Target with ID ${targetId} does not have a resource ID`
+                `Target with ID ${targetId} has invalid ownership`
            )
        );
    }

    try {
-        const resource = await db
-            .select()
-            .from(resources)
-            .where(eq(resources.resourceId, resourceId!))
-            .limit(1);
+        let orgId: string;

-        if (resource.length === 0) {
-            return next(
-                createHttpError(
-                    HttpCode.NOT_FOUND,
-                    `Resource with ID ${resourceId} not found`
-                )
-            );
-        }
+        if (resourceId) {
+            const [resource] = await db
+                .select()
+                .from(resources)
+                .where(eq(resources.resourceId, resourceId))
+                .limit(1);

-        if (!resource[0].orgId) {
-            return next(
-                createHttpError(
-                    HttpCode.INTERNAL_SERVER_ERROR,
-                    `resource with ID ${resourceId} does not have an organization ID`
-                )
-            );
+            if (!resource) {
+                return next(
+                    createHttpError(
+                        HttpCode.NOT_FOUND,
+                        `Resource with ID ${resourceId} not found`
+                    )
+                );
+            }
+
+            if (!resource.orgId) {
+                return next(
+                    createHttpError(
+                        HttpCode.INTERNAL_SERVER_ERROR,
+                        `Resource with ID ${resourceId} does not have an organization ID`
+                    )
+                );
+            }
+
+            orgId = resource.orgId;
+        } else {
+            const [provider] = await db
+                .select()
+                .from(aiProviders)
+                .where(eq(aiProviders.providerId, providerId!))
+                .limit(1);
+
+            if (!provider) {
+                return next(
+                    createHttpError(
+                        HttpCode.NOT_FOUND,
+                        `AI provider with ID ${providerId} not found`
+                    )
+                );
+            }
+
+            orgId = provider.orgId;
        }

        if (!req.userOrg) {
-            const res = await db
+            const userOrgResult = await db
                .select()
                .from(userOrgs)
                .where(
-                    and(
-                        eq(userOrgs.userId, userId),
-                        eq(userOrgs.orgId, resource[0].orgId)
-                    )
+                    and(eq(userOrgs.userId, userId), eq(userOrgs.orgId, orgId))
                );
-            req.userOrg = res[0];
+            req.userOrg = userOrgResult[0];
        }

-        if (!req.userOrg) {
-            next(
+        if (!req.userOrg || req.userOrg.orgId !== orgId) {
+            return next(
                createHttpError(
                    HttpCode.FORBIDDEN,
                    "User does not have access to this organization"
                )
            );
-        } else {
-            req.userOrgRoleIds = await getUserOrgRoleIds(
-                req.userOrg.userId,
-                resource[0].orgId!
-            );
-            req.userOrgId = resource[0].orgId!;
        }

-        const orgId = req.userOrg.orgId;
+        req.userOrgRoleIds = await getUserOrgRoleIds(req.userOrg.userId, orgId);
+        req.userOrgId = orgId;

-        if (req.orgPolicyAllowed === undefined && orgId) {
+        if (req.orgPolicyAllowed === undefined) {
            const policyCheck = await checkOrgAccessPolicy({
                orgId,
                userId,
@@ -128,22 +141,24 @@ export async function verifyTargetAccess(
            }
        }

-        const resourceAllowed = await canUserAccessResource({
-            userId,
-            resourceId,
-            roleIds: req.userOrgRoleIds ?? []
-        });
+        if (resourceId) {
+            const resourceAllowed = await canUserAccessResource({
+                userId,
+                resourceId,
+                roleIds: req.userOrgRoleIds ?? []
+            });

-        if (!resourceAllowed) {
-            return next(
-                createHttpError(
-                    HttpCode.FORBIDDEN,
-                    "User does not have access to this resource"
-                )
-            );
+            if (!resourceAllowed) {
+                return next(
+                    createHttpError(
+                        HttpCode.FORBIDDEN,
+                        "User does not have access to this resource"
+                    )
+                );
+            }
        }

-        next();
+        return next();
    } catch (e) {
        return next(
            createHttpError(
@@ -28,7 +28,9 @@ export enum OpenAPITags {
    HealthCheck = "Health Check",
    PublicResourcePolicyLegacy = "Public Resource Policy (Legacy)",
    PublicResourceLegacy = "Public Resource (Legacy)",
-    PrivateResourceLegacy = "Private Resource (Legacy)"
+    PrivateResourceLegacy = "Private Resource (Legacy)",
+    AiProvider = "AI Provider",
+    AiModel = "AI Model"
 }

 // Order here controls the order tags are displayed in Swagger UI
@@ -25,7 +25,6 @@ import {
    Transaction
 } from "@server/db";
 import logger from "@server/logger";
-import { ExitNodePingResult } from "@server/routers/newt";
 import { eq, and, or, ne, isNull, inArray } from "drizzle-orm";
 import axios from "axios";
 import config from "../config";
@@ -330,6 +329,16 @@ export async function listExitNodes(
    return exitNodesList;
 }

+export type ExitNodePingResult = {
+    exitNodeId: number;
+    latencyMs: number;
+    weight: number;
+    error?: string;
+    exitNodeName: string;
+    endpoint: string;
+    wasPreviouslyConnected: boolean;
+};
+
 /**
 * Selects the most suitable exit node from a list of ping results.
 *
@@ -18,6 +18,7 @@ import {
    domains,
    exitNodes,
    loginPage,
+    SiteResource,
    targetHealthCheck
 } from "@server/db";
 import {
@@ -40,7 +41,10 @@ import {
    siteNetworks,
    siteResources,
    Target,
-    targets
+    targets,
+    aiProviders,
+    resourceAiProviders,
+    siteResourceAiProviders
 } from "@server/db";
 import {
    sanitize,
@@ -54,6 +58,7 @@ import {
    getValidCertificatesForDomains
 } from "#private/lib/certificates";
 import { build } from "@server/build";
+import regionalCache from "#private/lib/cache";

 const redirectHttpsMiddlewareName = "redirect-to-https";
 const redirectToRootMiddlewareName = "redirect-to-root";
@@ -87,8 +92,24 @@ export async function getTraefikConfig(
    generateLoginPageRouters = false,
    allowRawResources = true,
    maintenancePageUiUrl: string | null = null,
-    browserGatewayUiUrl: string | null = null
+    browserGatewayUiUrl: string | null = null,
+    aiGatewayUrl: string | null = null
 ): Promise<any> {
+    // Get the exit node but cache it for 5 minutes to avoid hitting the DB too often
+    const exitNodeCacheKey = `exitNode:${exitNodeId}`;
+    let exitNode =
+        await regionalCache.get<typeof exitNodes.$inferSelect>(
+            exitNodeCacheKey
+        );
+    if (!exitNode) {
+        [exitNode] = await db
+            .select()
+            .from(exitNodes)
+            .where(eq(exitNodes.exitNodeId, exitNodeId))
+            .limit(1);
+        await regionalCache.set(exitNodeCacheKey, exitNode, 300);
+    }
+
    // Get resources with their targets and sites in a single optimized query
    // Start from sites on this exit node, then join to targets and resources
    const resourcesWithTargetsAndSites = await db
@@ -136,7 +157,7 @@ export async function getTraefikConfig(
            siteId: sites.siteId,
            siteType: sites.type,
            siteOnline: sites.online,
-            subnet: sites.subnet,
+            subnet: sites.exitNodeSubnet,
            exitNodeId: sites.exitNodeId,
            // Namespace
            domainNamespaceId: domainNamespaces.domainNamespaceId,
@@ -361,7 +382,7 @@ export async function getTraefikConfig(
    let siteResourcesWithFullDomain: {
        siteResourceId: number;
        fullDomain: string | null;
-        mode: "http" | "host" | "cidr" | "ssh";
+        mode: SiteResource["mode"];
    }[] = [];
    if (
        build == "enterprise" &&
@@ -386,7 +407,7 @@ export async function getTraefikConfig(
                and(
                    eq(siteResources.enabled, true),
                    isNotNull(siteResources.fullDomain),
-                    eq(siteResources.mode, "http"),
+                    eq(siteResources.mode, "http"), // important so we dont double get the inference siteResources below
                    eq(siteResources.ssl, true),
                    eq(sites.exitNodeId, exitNodeId),
                    inArray(sites.type, siteTypes)
@@ -394,6 +415,56 @@ export async function getTraefikConfig(
            );
    }

+    // Inference-mode resources/siteResources have no targets/sites/network
+    // (their "backend" is the central AI gateway, not something on a site),
+    // so they can't be reached via the joins above - query them separately
+    // and include them on every exit node.
+    const inferenceResources = await db
+        .selectDistinct({
+            resourceId: resources.resourceId,
+            fullDomain: resources.fullDomain,
+            ssl: resources.ssl,
+            subdomain: resources.subdomain,
+            domainId: resources.domainId,
+            enabled: resources.enabled,
+            wildcard: resources.wildcard,
+            domainCertResolver: domains.certResolver,
+            preferWildcardCert: domains.preferWildcardCert
+        })
+        .from(resources)
+        .innerJoin(
+            resourceAiProviders,
+            eq(resources.resourceId, resourceAiProviders.resourceId)
+        )
+        .innerJoin(
+            aiProviders,
+            eq(resourceAiProviders.providerId, aiProviders.providerId)
+        )
+        .leftJoin(domains, eq(domains.domainId, resources.domainId))
+        .where(
+            and(
+                eq(resources.mode, "inference"),
+                eq(resources.enabled, true),
+                eq(aiProviders.enabled, true)
+            )
+        );
+
+    const siteResourcesInference = await db
+        .selectDistinct({
+            siteResourceId: siteResources.siteResourceId,
+            fullDomain: siteResources.fullDomain,
+            ssl: siteResources.ssl,
+            enabled: siteResources.enabled
+        })
+        .from(siteResources)
+        .where(
+            and(
+                eq(siteResources.mode, "inference"),
+                eq(siteResources.enabled, true),
+                isNotNull(siteResources.fullDomain)
+            )
+        );
+
    let validCerts: CertificateResult[] = [];
    if (privateConfig.getRawPrivateConfig().flags.use_pangolin_dns) {
        // create a list of all domains to get certs for
@@ -415,6 +486,17 @@ export async function getTraefikConfig(
                domains.add(bgResource.fullDomain);
            }
        }
+        // Include inference resource/siteResource domains
+        for (const ir of inferenceResources) {
+            if (ir.enabled && ir.ssl && ir.fullDomain) {
+                domains.add(ir.fullDomain);
+            }
+        }
+        for (const sr of siteResourcesInference) {
+            if (sr.enabled && sr.ssl && sr.fullDomain) {
+                domains.add(sr.fullDomain);
+            }
+        }
        // get the valid certs for these domains
        validCerts = await getValidCertificatesForDomains(domains, true); // we are caching here because this is called often
        // logger.debug(`Valid certs for domains: ${JSON.stringify(validCerts)}`);
@@ -1461,6 +1543,244 @@ export async function getTraefikConfig(
        }
    }

+    if (aiGatewayUrl) {
+        // The AI gateway may live on a different host than the inference
+        // resource itself (e.g. a remote exit node forwarding to the
+        // central dashboard over a tunnel). passHostHeader would forward
+        // the resource's own Host, which that external host won't
+        // recognize, so we pin the Host header to the gateway's own host
+        // and smuggle the original resource host through in "p-host"
+        // instead (same pattern as the maintenance-page routes above).
+        let aiGatewayHost: string | undefined;
+        try {
+            aiGatewayHost = new URL(aiGatewayUrl).host;
+        } catch {
+            aiGatewayHost = undefined;
+        }
+
+        // The p-host smuggling above is only necessary when the AI gateway
+        // is overridden to a different host than the resource's own. In the
+        // default case, leave the Host header untouched so it's visible on
+        // the other end.
+        const aiGatewayOverride =
+            config.getRawConfig().server.ai_gateway_override;
+
+        // Public inference resources: same TLS/cert-resolver handling as
+        // plain http-mode resources, but the service points at the AI
+        // gateway instead of any real backend targets.
+        for (const ir of inferenceResources) {
+            if (!ir.enabled) continue;
+            if (!ir.domainId || !ir.fullDomain) continue;
+
+            if (!config_output.http.routers) config_output.http.routers = {};
+            if (!config_output.http.services) config_output.http.services = {};
+
+            const fullDomain = ir.fullDomain;
+            const irKey = `inference-r${ir.resourceId}`;
+            const routerName = `${irKey}-router`;
+            const serviceName = `${irKey}-service`;
+
+            let rule: string;
+            if (ir.wildcard && fullDomain.startsWith("*.")) {
+                const escaped = fullDomain.slice(2).replace(/\./g, "\\.");
+                rule = `HostRegexp(\`^[^.]+\\.${escaped}$\`)`;
+            } else {
+                rule = `Host(\`${fullDomain}\`)`;
+            }
+
+            let tls: any = {};
+            if (!privateConfig.getRawPrivateConfig().flags.use_pangolin_dns) {
+                const domainParts = fullDomain.split(".");
+                let wildCard;
+                if (domainParts.length <= 2) {
+                    wildCard = `*.${domainParts.join(".")}`;
+                } else {
+                    wildCard = `*.${domainParts.slice(1).join(".")}`;
+                }
+                if (!ir.subdomain) {
+                    wildCard = fullDomain;
+                }
+
+                const globalDefaultResolver =
+                    config.getRawConfig().traefik.cert_resolver;
+                const globalDefaultPreferWildcard =
+                    config.getRawConfig().traefik.prefer_wildcard_cert;
+                const resolverName = ir.domainCertResolver
+                    ? ir.domainCertResolver.trim()
+                    : globalDefaultResolver;
+                const preferWildcard =
+                    ir.preferWildcardCert !== undefined &&
+                    ir.preferWildcardCert !== null
+                        ? ir.preferWildcardCert
+                        : globalDefaultPreferWildcard;
+
+                tls = {
+                    certResolver: resolverName,
+                    ...(preferWildcard ? { domains: [{ main: wildCard }] } : {})
+                };
+            } else {
+                const matchingCert = validCerts.find(
+                    (cert) => cert.queriedDomain === fullDomain
+                );
+                if (!matchingCert) {
+                    logger.debug(
+                        `No matching certificate found for inference resource domain: ${fullDomain}`
+                    );
+                    continue;
+                }
+            }
+
+            const additionalMiddlewares =
+                config.getRawConfig().traefik.additional_middlewares || [];
+            const routerMiddlewares = [badgerMiddlewareName];
+
+            if (aiGatewayOverride) {
+                const irHeadersMiddlewareName = `${irKey}-headers-middleware`;
+                config_output.http.middlewares[irHeadersMiddlewareName] = {
+                    headers: {
+                        customRequestHeaders: {
+                            ...(aiGatewayHost ? { Host: aiGatewayHost } : {}),
+                            "p-host": fullDomain
+                        }
+                    }
+                };
+                routerMiddlewares.push(irHeadersMiddlewareName);
+            }
+
+            routerMiddlewares.push(...additionalMiddlewares);
+
+            if (ir.ssl) {
+                config_output.http.routers[routerName + "-redirect"] = {
+                    entryPoints: [
+                        config.getRawConfig().traefik.http_entrypoint
+                    ],
+                    middlewares: [redirectHttpsMiddlewareName],
+                    service: serviceName,
+                    rule,
+                    priority: 100
+                };
+            }
+
+            config_output.http.routers[routerName] = {
+                entryPoints: [
+                    ir.ssl
+                        ? config.getRawConfig().traefik.https_entrypoint
+                        : config.getRawConfig().traefik.http_entrypoint
+                ],
+                middlewares: routerMiddlewares,
+                service: serviceName,
+                rule,
+                priority: 100,
+                ...(ir.ssl ? { tls } : {})
+            };
+
+            config_output.http.services[serviceName] = {
+                loadBalancer: {
+                    servers: [{ url: aiGatewayUrl }]
+                }
+            };
+        }
+
+        // Private (siteResource) inference resources: routed by their alias
+        // instead of a public fullDomain, and deliberately WITHOUT the
+        // badger middleware - no per-user auth/policy stack exists for
+        // siteResources today (see plan doc), so gating here is
+        // reachability-only for now.
+        for (const sr of siteResourcesInference) {
+            if (!sr.enabled || !sr.fullDomain) continue;
+
+            if (!config_output.http.routers) config_output.http.routers = {};
+            if (!config_output.http.services) config_output.http.services = {};
+
+            const fullDomain = sr.fullDomain;
+            const srKey = `inference-sr${sr.siteResourceId}`;
+            const routerName = `${srKey}-router`;
+            const serviceName = `${srKey}-service`;
+            const rule = `Host(\`${fullDomain}\`) && ClientIP(\`${exitNode.address}\`)`; // restrict to coming from the exit node ip range that the client is connected to
+
+            let tls: any = {};
+            if (!privateConfig.getRawPrivateConfig().flags.use_pangolin_dns) {
+                const domainParts = fullDomain.split(".");
+                const wildCard =
+                    domainParts.length <= 2
+                        ? `*.${domainParts.join(".")}`
+                        : `*.${domainParts.slice(1).join(".")}`;
+
+                const globalDefaultResolver =
+                    config.getRawConfig().traefik.cert_resolver;
+                const globalDefaultPreferWildcard =
+                    config.getRawConfig().traefik.prefer_wildcard_cert;
+
+                tls = {
+                    certResolver: globalDefaultResolver,
+                    ...(globalDefaultPreferWildcard
+                        ? { domains: [{ main: wildCard }] }
+                        : {})
+                };
+            } else {
+                const matchingCert = validCerts.find(
+                    (cert) => cert.queriedDomain === fullDomain
+                );
+                if (!matchingCert) {
+                    logger.debug(
+                        `No matching certificate found for inference siteResource fullDomain: ${fullDomain}`
+                    );
+                    continue;
+                }
+            }
+
+            const additionalMiddlewares =
+                config.getRawConfig().traefik.additional_middlewares || [];
+            const routerMiddlewares: string[] = [];
+
+            if (aiGatewayOverride) {
+                const srHeadersMiddlewareName = `${srKey}-headers-middleware`;
+                config_output.http.middlewares[srHeadersMiddlewareName] = {
+                    headers: {
+                        customRequestHeaders: {
+                            ...(aiGatewayHost ? { Host: aiGatewayHost } : {}),
+                            "p-host": fullDomain
+                        }
+                    }
+                };
+                routerMiddlewares.push(srHeadersMiddlewareName);
+            }
+
+            routerMiddlewares.push(...additionalMiddlewares);
+
+            if (sr.ssl) {
+                config_output.http.routers[routerName + "-redirect"] = {
+                    entryPoints: [
+                        config.getRawConfig().traefik.http_entrypoint
+                    ],
+                    middlewares: [redirectHttpsMiddlewareName],
+                    service: serviceName,
+                    rule,
+                    priority: 100
+                };
+            }
+
+            config_output.http.routers[routerName] = {
+                entryPoints: [
+                    sr.ssl
+                        ? config.getRawConfig().traefik.https_entrypoint
+                        : config.getRawConfig().traefik.http_entrypoint
+                ],
+                middlewares: routerMiddlewares,
+                service: serviceName,
+                rule,
+                priority: 100,
+                ...(sr.ssl ? { tls } : {})
+            };
+
+            config_output.http.services[serviceName] = {
+                loadBalancer: {
+                    servers: [{ url: aiGatewayUrl }]
+                }
+            };
+        }
+    }
+
    if (generateLoginPageRouters) {
        const exitNodeLoginPages = await db
            .select({
@@ -10,12 +10,12 @@
 *
 * This file is not licensed under the AGPLv3.
 */
-import { certificates, db, domains, orgDomains } from "@server/db";
+import { certificates, db, domainNamespaces, domains, orgDomains } from "@server/db";
 import response from "@server/lib/response";
 import logger from "@server/logger";
 import { type GetBatchedCertificateResponse } from "@server/routers/certificates/types";
 import HttpCode from "@server/types/HttpCode";
-import { and, eq, inArray, or } from "drizzle-orm";
+import { and, eq, inArray, isNotNull, or } from "drizzle-orm";
 import { NextFunction, Request, Response } from "express";
 import createHttpError from "http-errors";
 import { z } from "zod";
@@ -63,14 +63,28 @@ async function query(orgId: string, domainList: string[]) {
        })
        .from(certificates)
        .innerJoin(domains, eq(certificates.domainId, domains.domainId))
-        .innerJoin(
+        .leftJoin(
            orgDomains,
            and(
                eq(domains.domainId, orgDomains.domainId),
                eq(orgDomains.orgId, orgId)
            )
        )
-        .where(and(inArray(certificates.domain, domainList)));
+        .leftJoin(
+            domainNamespaces,
+            eq(domains.domainId, domainNamespaces.domainId)
+        )
+        .where(
+            and(
+                inArray(certificates.domain, domainList),
+                // Namespace domains are shared across all orgs, so they skip
+                // the org-ownership check (mirrors verifyCertificateAccess).
+                or(
+                    isNotNull(orgDomains.orgId),
+                    isNotNull(domainNamespaces.domainNamespaceId)
+                )
+            )
+        );

    // All non resolved domain certificates might be `ns` or `wildcard`,
    // which means exact domain certificates do not exist
@@ -110,19 +124,27 @@ async function query(orgId: string, domainList: string[]) {
            })
            .from(certificates)
            .innerJoin(domains, eq(certificates.domainId, domains.domainId))
-            .innerJoin(
+            .leftJoin(
                orgDomains,
                and(
                    eq(domains.domainId, orgDomains.domainId),
                    eq(orgDomains.orgId, orgId)
                )
            )
+            .leftJoin(
+                domainNamespaces,
+                eq(domains.domainId, domainNamespaces.domainId)
+            )
            .where(
                and(
                    eq(certificates.wildcard, true),
                    or(
                        inArray(certificates.domain, [...domainLevelDownSet]),
                        inArray(certificates.domain, [...wildcardDomainSet])
+                    ),
+                    or(
+                        isNotNull(orgDomains.orgId),
+                        isNotNull(domainNamespaces.domainNamespaceId)
                    )
                )
            );
@@ -351,6 +351,7 @@ hybridRouter.get(
        }

        const pangolinUIUrl = config.getRawConfig().app.dashboard_url; // points to the dashboard to serve from there
+        const aiGatewayUrl = `${config.getRawConfig().app.dashboard_url}/api/v1/ai-gateway`;

        try {
            const traefikConfig = await getTraefikConfig(
@@ -360,7 +361,8 @@ hybridRouter.get(
                false, // Dont include login pages,
                true, // allow raw resources
                pangolinUIUrl, // dont generate maintenance page
-                pangolinUIUrl // generate browser gateway targets
+                pangolinUIUrl, // generate browser gateway targets
+                aiGatewayUrl
            );

            return response(res, {
@@ -178,7 +178,7 @@ export async function reGenerateSiteSecret(
                    );
                }

-                if (site.exitNodeId && site.subnet) {
+                if (site.exitNodeId && site.exitNodeSubnet) {
                    await deletePeer(site.exitNodeId, site.pubKey!); // the old pubkey
                    await addPeer(site.exitNodeId, {
                        publicKey: pubKey,
@@ -0,0 +1,21 @@
+import { Router } from "express";
+import {
+    AI_CAPABILITY_DEFS,
+    type AiCapability
+} from "@server/lib/aiCapabilities";
+import { handleAiGatewayProxy } from "@server/routers/aiGateway/pipeline";
+
+export function createAiGatewayRouter() {
+    const router = Router();
+
+    for (const def of Object.values(AI_CAPABILITY_DEFS)) {
+        const capability = def.id as AiCapability;
+        for (const route of def.routes) {
+            router.post(route.path, (req, res) =>
+                handleAiGatewayProxy(req, res, capability)
+            );
+        }
+    }
+
+    return router;
+}
@@ -0,0 +1,2 @@
+export { handleAiGatewayProxy } from "./pipeline";
+export { createAiGatewayRouter } from "./createAiGatewayRouter";
@@ -0,0 +1,841 @@
+import { Request, Response } from "express";
+import { and, eq, inArray } from "drizzle-orm";
+import {
+    AiProvider,
+    aiModels,
+    aiProviders,
+    clients,
+    db,
+    exitNodes,
+    resourceAiModels,
+    resourceAiProviders,
+    resources,
+    siteResourceAiModels,
+    siteResourceAiProviders,
+    siteResources,
+    users
+} from "@server/db";
+import config from "@server/lib/config";
+import { decrypt } from "@server/lib/crypto";
+import {
+    AiProviderAuthType,
+    AiProviderType,
+    applyAiProviderAuthHeaders,
+    applyAiProviderCustomHeaders,
+    authTypeRequiresApiKey
+} from "@server/lib/aiProviderDefaults";
+import {
+    AI_CAPABILITY_DEFS,
+    providerHasCapability,
+    type AiCapability
+} from "@server/lib/aiCapabilities";
+import { proxyAiGatewayToSiteTarget } from "@server/routers/aiGateway/targetRouting";
+import {
+    SESSION_COOKIE_NAME,
+    validateSessionToken
+} from "@server/auth/sessions/app";
+import { getUserOrgRoles } from "@server/lib/userOrgRoles";
+import { isIpInCidr } from "@server/lib/ip";
+import { localCache } from "@server/lib/cache";
+import logger from "@server/logger";
+import HttpCode from "@server/types/HttpCode";
+import {
+    resolveEffectiveLists,
+    type AccessMode,
+    type ModelListType
+} from "@server/lib/aiInferenceResource";
+import {
+    compareModelKeySpecificity,
+    isAllowedByLists,
+    mostSpecificMatchingAllow
+} from "@server/lib/aiModelKeyMatch";
+import { aiGatewayUpstreamFetch } from "@server/lib/aiGatewayUpstreamFetch";
+import { getModelPricing, calculateAiCost } from "@server/lib/aiModelPricing";
+import {
+    extractUsage,
+    estimateUsage,
+    isUsageEmpty,
+    needsStreamUsageInjection,
+    withStreamUsageOption,
+    stripInjectedUsageFrame,
+    extractResponseModel,
+    type AiUsage
+} from "@server/lib/aiUsageExtraction";
+
+const EXIT_NODE_RANGES_CACHE_KEY = "aiGateway:exitNodeRanges";
+const EXIT_NODE_RANGES_TTL_SEC = 6000;
+const CLIENT_BY_IP_TTL_SEC = 30;
+const REQUEST_USER_TTL_SEC = 30;
+
+type CachedClient = { clientId: number; userId: string | null } | null;
+
+async function getExitNodeRanges(): Promise<string[]> {
+    const cached = localCache.get<string[]>(EXIT_NODE_RANGES_CACHE_KEY);
+    if (cached) {
+        return cached;
+    }
+
+    const rows = await db
+        .select({ address: exitNodes.address })
+        .from(exitNodes);
+    const ranges = rows.map((r) => r.address);
+
+    localCache.set(
+        EXIT_NODE_RANGES_CACHE_KEY,
+        ranges,
+        EXIT_NODE_RANGES_TTL_SEC
+    );
+    return ranges;
+}
+
+async function findClientByIp(ip: string): Promise<CachedClient> {
+    const cacheKey = `aiGateway:clientByIp:${ip}`;
+    const cached = localCache.get<CachedClient>(cacheKey);
+    if (cached !== undefined) {
+        return cached;
+    }
+
+    const [client] = await db
+        .select({ clientId: clients.clientId, userId: clients.userId })
+        .from(clients)
+        .where(eq(clients.exitNodeSubnet, `${ip}/32`))
+        .limit(1);
+
+    const result: CachedClient = client || null;
+    localCache.set(cacheKey, result, CLIENT_BY_IP_TTL_SEC);
+    return result;
+}
+
+type ProviderAttachment = {
+    provider: AiProvider;
+    accessMode: AccessMode;
+};
+
+type ResourceModelPattern = {
+    providerId: number;
+    modelKey: string;
+    listType: ModelListType;
+    enabled: boolean;
+};
+
+type ProviderPatternLists = {
+    allows: string[];
+    blocks: string[];
+};
+
+type ResolvedTarget = {
+    resourceId: number | null;
+    siteResourceId: number | null;
+    orgId: string | null;
+    attachments: ProviderAttachment[];
+    resourceListsByProvider: Map<number, ProviderPatternLists>;
+};
+
+type ProviderSelection =
+    | { ok: true; provider: AiProvider }
+    | { ok: false; status: number; message: string };
+
+export type RequestUser = {
+    userId: string;
+    username: string;
+    email: string | null;
+    name: string | null;
+    role: string | null;
+};
+
+// Identity headers forwarded to the upstream inference endpoint when the
+// requesting user is known. Omitted entirely (not sent empty) when we
+// couldn't resolve a user for the request.
+export function applyRequestUserHeaders(
+    headers: Record<string, string>,
+    requestUser: RequestUser | null
+): void {
+    if (!requestUser) {
+        return;
+    }
+    headers["Remote-User"] = requestUser.username;
+    if (requestUser.email) {
+        headers["Remote-Email"] = requestUser.email;
+    }
+    if (requestUser.name) {
+        headers["Remote-Name"] = requestUser.name;
+    }
+    if (requestUser.role) {
+        headers["Remote-Role"] = requestUser.role;
+    }
+}
+
+async function buildRequestUser(
+    userId: string,
+    orgId: string | null
+): Promise<RequestUser | null> {
+    const cacheKey = `aiGateway:requestUser:${userId}:${orgId || ""}`;
+    const cached = localCache.get<RequestUser | null>(cacheKey);
+    if (cached !== undefined) {
+        return cached;
+    }
+
+    const [user] = await db
+        .select()
+        .from(users)
+        .where(eq(users.userId, userId))
+        .limit(1);
+
+    if (!user) {
+        localCache.set(cacheKey, null, REQUEST_USER_TTL_SEC);
+        return null;
+    }
+
+    const orgRoles = orgId ? await getUserOrgRoles(user.userId, orgId) : [];
+
+    const requestUser: RequestUser = {
+        userId: user.userId,
+        username: user.username,
+        email: user.email,
+        name: user.name,
+        role: orgRoles.map((r) => r.roleName).join(", ") || null
+    };
+
+    localCache.set(cacheKey, requestUser, REQUEST_USER_TTL_SEC);
+    return requestUser;
+}
+
+async function resolveRequestUser(
+    req: Request,
+    _resourceId: number | null,
+    orgId: string | null
+): Promise<RequestUser | null> {
+    const sessionToken = req.cookies?.[SESSION_COOKIE_NAME];
+    if (sessionToken) {
+        const { session, user } = await validateSessionToken(sessionToken);
+        if (session && user) {
+            return buildRequestUser(user.userId, orgId);
+        }
+    }
+
+    // TODO: MAKE SURE THIS CAN NOT BE SPOOFED AND CAN BE TRUSTED AS AN INTERNAL ADDRESS FROM A NODE
+
+    const ip = req.ip;
+    if (!ip) {
+        return null;
+    }
+
+    const exitNodeRanges = await getExitNodeRanges();
+    const inExitNodeRange = exitNodeRanges.some((range) =>
+        isIpInCidr(ip, range)
+    );
+    if (!inExitNodeRange) {
+        return null;
+    }
+
+    const client = await findClientByIp(ip);
+    if (!client || !client.userId) {
+        return null;
+    }
+
+    return buildRequestUser(client.userId, orgId);
+}
+
+async function resolveTarget(host: string): Promise<ResolvedTarget | null> {
+    const [[resourceRow], [siteResourceRow]] = await Promise.all([
+        db
+            .select({
+                resourceId: resources.resourceId,
+                orgId: resources.orgId
+            })
+            .from(resources)
+            .where(
+                and(
+                    eq(resources.fullDomain, host),
+                    eq(resources.mode, "inference"),
+                    eq(resources.enabled, true)
+                )
+            )
+            .limit(1),
+        db
+            .select({
+                siteResourceId: siteResources.siteResourceId,
+                orgId: siteResources.orgId
+            })
+            .from(siteResources)
+            .where(
+                and(
+                    eq(siteResources.fullDomain, host),
+                    eq(siteResources.mode, "inference"),
+                    eq(siteResources.enabled, true)
+                )
+            )
+            .limit(1)
+    ]);
+
+    // Prefer public inference resources when both match the same host.
+    if (resourceRow) {
+        const [attachmentRows, resourcePatterns] = await Promise.all([
+            db
+                .select({
+                    provider: aiProviders,
+                    accessMode: resourceAiProviders.accessMode
+                })
+                .from(resourceAiProviders)
+                .innerJoin(
+                    aiProviders,
+                    eq(resourceAiProviders.providerId, aiProviders.providerId)
+                )
+                .where(
+                    and(
+                        eq(
+                            resourceAiProviders.resourceId,
+                            resourceRow.resourceId
+                        ),
+                        eq(aiProviders.enabled, true),
+                        eq(resourceAiProviders.enabled, true)
+                    )
+                ),
+            db
+                .select({
+                    providerId: aiModels.providerId,
+                    modelKey: aiModels.modelKey,
+                    listType: resourceAiModels.listType,
+                    enabled: aiModels.enabled
+                })
+                .from(resourceAiModels)
+                .innerJoin(
+                    aiModels,
+                    eq(resourceAiModels.modelId, aiModels.modelId)
+                )
+                .where(eq(resourceAiModels.resourceId, resourceRow.resourceId))
+        ]);
+
+        if (attachmentRows.length === 0) {
+            return null;
+        }
+
+        return {
+            resourceId: resourceRow.resourceId,
+            siteResourceId: null,
+            orgId: resourceRow.orgId,
+            attachments: attachmentRows.map((a) => ({
+                provider: a.provider,
+                accessMode: a.accessMode
+            })),
+            resourceListsByProvider: groupPatternsByProvider(resourcePatterns)
+        };
+    }
+
+    if (siteResourceRow) {
+        const [attachmentRows, resourcePatterns] = await Promise.all([
+            db
+                .select({
+                    provider: aiProviders,
+                    accessMode: siteResourceAiProviders.accessMode
+                })
+                .from(siteResourceAiProviders)
+                .innerJoin(
+                    aiProviders,
+                    eq(
+                        siteResourceAiProviders.providerId,
+                        aiProviders.providerId
+                    )
+                )
+                .where(
+                    and(
+                        eq(
+                            siteResourceAiProviders.siteResourceId,
+                            siteResourceRow.siteResourceId
+                        ),
+                        eq(aiProviders.enabled, true),
+                        eq(siteResourceAiProviders.enabled, true)
+                    )
+                ),
+            db
+                .select({
+                    providerId: aiModels.providerId,
+                    modelKey: aiModels.modelKey,
+                    listType: siteResourceAiModels.listType,
+                    enabled: aiModels.enabled
+                })
+                .from(siteResourceAiModels)
+                .innerJoin(
+                    aiModels,
+                    eq(siteResourceAiModels.modelId, aiModels.modelId)
+                )
+                .where(
+                    eq(
+                        siteResourceAiModels.siteResourceId,
+                        siteResourceRow.siteResourceId
+                    )
+                )
+        ]);
+
+        if (attachmentRows.length === 0) {
+            return null;
+        }
+
+        return {
+            resourceId: null,
+            siteResourceId: siteResourceRow.siteResourceId,
+            orgId: siteResourceRow.orgId,
+            attachments: attachmentRows.map((a) => ({
+                provider: a.provider,
+                accessMode: a.accessMode
+            })),
+            resourceListsByProvider: groupPatternsByProvider(resourcePatterns)
+        };
+    }
+
+    return null;
+}
+
+function groupPatternsByProvider(
+    patterns: ResourceModelPattern[]
+): Map<number, ProviderPatternLists> {
+    const byProvider = new Map<number, ProviderPatternLists>();
+    for (const pattern of patterns) {
+        if (!pattern.enabled) {
+            continue;
+        }
+        let lists = byProvider.get(pattern.providerId);
+        if (!lists) {
+            lists = { allows: [], blocks: [] };
+            byProvider.set(pattern.providerId, lists);
+        }
+        if (pattern.listType === "allow") {
+            lists.allows.push(pattern.modelKey);
+        } else {
+            lists.blocks.push(pattern.modelKey);
+        }
+    }
+    return byProvider;
+}
+
+async function selectProvider(
+    attachments: ProviderAttachment[],
+    resourceListsByProvider: Map<number, ProviderPatternLists>,
+    requestedModel: string | undefined
+): Promise<ProviderSelection> {
+    if (!requestedModel) {
+        return {
+            ok: false,
+            status: HttpCode.FORBIDDEN,
+            message: "A model must be specified for this resource"
+        };
+    }
+
+    const attachmentByProviderId = new Map(
+        attachments.map((a) => [a.provider.providerId, a])
+    );
+    const providerIds = [...attachmentByProviderId.keys()];
+    if (providerIds.length === 0) {
+        return {
+            ok: false,
+            status: HttpCode.FORBIDDEN,
+            message: `Model "${requestedModel}" is not permitted on this resource`
+        };
+    }
+
+    const providerModels = await db
+        .select({
+            providerId: aiModels.providerId,
+            modelKey: aiModels.modelKey,
+            listType: aiModels.listType,
+            enabled: aiModels.enabled
+        })
+        .from(aiModels)
+        .where(inArray(aiModels.providerId, providerIds));
+
+    const allowsByProvider = new Map<number, string[]>();
+    const blocksByProvider = new Map<number, string[]>();
+    for (const model of providerModels) {
+        if (!model.enabled) {
+            continue;
+        }
+        const targetMap =
+            model.listType === "allow" ? allowsByProvider : blocksByProvider;
+        const existing = targetMap.get(model.providerId) ?? [];
+        existing.push(model.modelKey);
+        targetMap.set(model.providerId, existing);
+    }
+
+    type ModelCandidate = {
+        provider: AiProvider;
+        modelKey: string;
+    };
+
+    const candidates: ModelCandidate[] = [];
+    for (const [providerId, attachment] of attachmentByProviderId) {
+        const resourceLists = resourceListsByProvider.get(providerId);
+        const { allows, blocks } = resolveEffectiveLists({
+            accessMode: attachment.accessMode,
+            providerAllows: allowsByProvider.get(providerId) ?? [],
+            providerBlocks: blocksByProvider.get(providerId) ?? [],
+            resourceAllows: resourceLists?.allows ?? [],
+            resourceBlocks: resourceLists?.blocks ?? []
+        });
+
+        if (!isAllowedByLists(requestedModel, allows, blocks)) {
+            continue;
+        }
+        const matchingAllow = mostSpecificMatchingAllow(requestedModel, allows);
+        if (!matchingAllow) {
+            continue;
+        }
+        candidates.push({
+            provider: attachment.provider,
+            modelKey: matchingAllow
+        });
+    }
+
+    if (candidates.length === 0) {
+        return {
+            ok: false,
+            status: HttpCode.FORBIDDEN,
+            message: `Model "${requestedModel}" is not permitted on this resource`
+        };
+    }
+
+    candidates.sort((a, b) =>
+        compareModelKeySpecificity(a.modelKey, b.modelKey)
+    );
+
+    const bestSpecificity = candidates[0].modelKey;
+    const topCandidates = candidates.filter(
+        (c) => compareModelKeySpecificity(c.modelKey, bestSpecificity) === 0
+    );
+
+    const uniqueProviders = new Map<number, AiProvider>();
+    for (const candidate of topCandidates) {
+        uniqueProviders.set(candidate.provider.providerId, candidate.provider);
+    }
+
+    if (uniqueProviders.size === 1) {
+        return { ok: true, provider: [...uniqueProviders.values()][0] };
+    }
+
+    return {
+        ok: false,
+        status: HttpCode.FORBIDDEN,
+        message: `Model "${requestedModel}" is ambiguous across multiple AI providers on this resource`
+    };
+}
+
+function logAiUsageAndCost(args: {
+    capability: AiCapability;
+    provider: AiProvider;
+    requestedModel: string | undefined;
+    requestBody: unknown;
+    responseText: string;
+    isStream: boolean;
+    headers: Headers;
+}): void {
+    const {
+        capability,
+        provider,
+        requestedModel,
+        requestBody,
+        responseText,
+        isStream,
+        headers
+    } = args;
+
+    let usage: AiUsage | null = extractUsage(
+        capability,
+        responseText,
+        isStream,
+        headers
+    );
+    if (!usage || isUsageEmpty(usage)) {
+        usage = estimateUsage(JSON.stringify(requestBody ?? ""), responseText);
+    }
+
+    const model = extractResponseModel(responseText) ?? requestedModel;
+    const pricing = getModelPricing(provider.type as AiProviderType, model);
+    const cost = calculateAiCost(pricing, usage);
+
+    logger.info("AI gateway request usage", {
+        capability,
+        providerId: provider.providerId,
+        providerType: provider.type,
+        model,
+        estimated: usage.estimated,
+        promptTokens: usage.promptTokens,
+        cacheReadTokens: usage.cacheReadTokens,
+        cacheWriteTokens: usage.cacheWriteTokens,
+        completionTokens: usage.completionTokens,
+        reasoningTokens: usage.reasoningTokens,
+        pricingApproximate: pricing?.approximate ?? null,
+        totalCostUsd: cost?.totalCost ?? null
+    });
+}
+
+export async function handleAiGatewayProxy(
+    req: Request,
+    res: Response,
+    capability: AiCapability
+): Promise<any> {
+    try {
+        const def = AI_CAPABILITY_DEFS[capability];
+
+        const host = (
+            (req.headers["p-host"] as string | undefined) || // p-host is only used sometimes when overriding the host header for some middleware proxy
+            req.headers.host ||
+            ""
+        ).split(":")[0];
+        if (!host) {
+            return res
+                .status(HttpCode.BAD_REQUEST)
+                .json({ error: { message: "Missing Host header" } });
+        }
+
+        logger.info(`AI gateway ${capability} request for host: ${host}`);
+
+        const target = await resolveTarget(host);
+        if (!target) {
+            return res.status(HttpCode.NOT_FOUND).json({
+                error: {
+                    message: "No inference resource found for this host"
+                }
+            });
+        }
+
+        const { attachments, resourceListsByProvider, resourceId, orgId } =
+            target;
+
+        const capableAttachments = attachments.filter((a) =>
+            providerHasCapability(a.provider.capabilities, capability)
+        );
+
+        if (capableAttachments.length === 0) {
+            return res.status(HttpCode.FORBIDDEN).json({
+                error: {
+                    message: `No AI provider on this resource supports ${capability}`
+                }
+            });
+        }
+
+        const requestedModel = def.extractModel(req);
+
+        const [requestUser, selection] = await Promise.all([
+            resolveRequestUser(req, resourceId, orgId),
+            selectProvider(
+                capableAttachments,
+                resourceListsByProvider,
+                requestedModel
+            )
+        ]);
+
+        if (requestUser) {
+            logger.debug(
+                `AI gateway request from user ${requestUser.userId} (${requestUser.username})`
+            );
+        }
+
+        if (!selection.ok) {
+            return res.status(selection.status).json({
+                error: { message: selection.message }
+            });
+        }
+
+        const { provider } = selection;
+
+        if (provider.type === "custom" && provider.routingMode === "target") {
+            return await proxyAiGatewayToSiteTarget(
+                req,
+                res,
+                provider,
+                requestUser,
+                capability
+            );
+        }
+
+        const upstreamUrl = provider.upstreamUrl;
+        const authType = provider.authType as AiProviderAuthType;
+
+        if (!upstreamUrl) {
+            return res.status(HttpCode.INTERNAL_SERVER_ERROR).json({
+                error: {
+                    message: "AI provider has no upstream URL configured"
+                }
+            });
+        }
+
+        let apiKey: string | null = null;
+        if (authTypeRequiresApiKey(authType)) {
+            if (!provider.apiKey) {
+                return res.status(HttpCode.INTERNAL_SERVER_ERROR).json({
+                    error: {
+                        message: "AI provider has no API key configured"
+                    }
+                });
+            }
+            const secret = config.getRawConfig().server.secret!;
+            apiKey = decrypt(provider.apiKey, secret);
+        }
+
+        const targetUrl = def.resolveUpstreamUrl(
+            upstreamUrl,
+            req,
+            requestedModel!
+        );
+
+        const skipHeaders = new Set([
+            "p-host",
+            "host",
+            "connection",
+            "keep-alive",
+            "proxy-authenticate",
+            "proxy-authorization",
+            "te",
+            "trailers",
+            "transfer-encoding",
+            "upgrade",
+            "content-length",
+            "accept-encoding"
+        ]);
+
+        const headers: Record<string, string> = {};
+        for (const [key, value] of Object.entries(req.headers)) {
+            if (skipHeaders.has(key.toLowerCase()) || value === undefined) {
+                continue;
+            }
+            headers[key] = Array.isArray(value) ? value.join(", ") : value;
+        }
+        applyAiProviderCustomHeaders(
+            headers,
+            provider.headers,
+            config.getRawConfig().server.secret!
+        );
+        applyAiProviderAuthHeaders(headers, authType, apiKey);
+        applyRequestUserHeaders(headers, requestUser);
+
+        // OpenAI's Chat Completions API only reports usage in a streaming
+        // response when asked to via stream_options.include_usage - inject
+        // it ourselves when the caller didn't, so we can still track cost,
+        // and strip the extra frame it adds back out of what we forward.
+        const injectedUsageOurselves = needsStreamUsageInjection(
+            capability,
+            req.body
+        );
+        const outboundBody = injectedUsageOurselves
+            ? withStreamUsageOption(req.body)
+            : req.body;
+        const body = JSON.stringify(outboundBody);
+
+        logger.debug("AI gateway upstream request", {
+            capability,
+            url: targetUrl,
+            method: "POST",
+            headers,
+            body: outboundBody,
+            skipTlsVerification: provider.skipTlsVerification
+        });
+
+        const abortController = new AbortController();
+        const onClientClose = () => {
+            if (!res.writableEnded) {
+                abortController.abort();
+            }
+        };
+        res.on("close", onClientClose);
+
+        let upstreamRes: globalThis.Response;
+        try {
+            upstreamRes = await aiGatewayUpstreamFetch(targetUrl, {
+                method: "POST",
+                headers,
+                body,
+                skipTlsVerification: provider.skipTlsVerification,
+                signal: abortController.signal
+            });
+        } catch (fetchError) {
+            res.off("close", onClientClose);
+            if (abortController.signal.aborted) {
+                // Client already disconnected; nothing left to respond to.
+                return;
+            }
+            logger.error({
+                message: "AI gateway upstream fetch failed",
+                url: targetUrl,
+                error: fetchError,
+                cause:
+                    fetchError instanceof Error
+                        ? (fetchError as Error & { cause?: unknown }).cause
+                        : undefined
+            });
+            throw fetchError;
+        }
+
+        const contentType = upstreamRes.headers.get("content-type") || "";
+        const isStream = def.isStreaming(req, contentType);
+
+        res.status(upstreamRes.status);
+        res.setHeader("Content-Type", contentType || "application/json");
+
+        if (isStream && upstreamRes.body) {
+            res.flushHeaders();
+            const reader = upstreamRes.body.getReader();
+            const decoder = new TextDecoder();
+            let fullText = "";
+            // Frame-boundary buffer, only used when we need to filter the
+            // usage-only frame we injected out of what reaches the client.
+            let sseCarry = "";
+            try {
+                while (!abortController.signal.aborted) {
+                    const { done, value } = await reader.read();
+                    if (done) break;
+                    const chunkText = decoder.decode(value, { stream: true });
+                    fullText += chunkText;
+                    if (injectedUsageOurselves) {
+                        sseCarry += chunkText;
+                        const lastBoundary = sseCarry.lastIndexOf("\n\n");
+                        if (lastBoundary !== -1) {
+                            const toEmit = sseCarry.slice(0, lastBoundary + 2);
+                            sseCarry = sseCarry.slice(lastBoundary + 2);
+                            res.write(stripInjectedUsageFrame(toEmit));
+                        }
+                    } else {
+                        res.write(value);
+                    }
+                }
+                if (injectedUsageOurselves && sseCarry) {
+                    res.write(stripInjectedUsageFrame(sseCarry));
+                }
+            } finally {
+                await reader.cancel().catch(() => {});
+                res.off("close", onClientClose);
+            }
+            if (!res.writableEnded) {
+                res.end();
+            }
+            if (!abortController.signal.aborted) {
+                logAiUsageAndCost({
+                    capability,
+                    provider,
+                    requestedModel,
+                    requestBody: req.body,
+                    responseText: fullText,
+                    isStream: true,
+                    headers: upstreamRes.headers
+                });
+            }
+            return;
+        }
+
+        res.off("close", onClientClose);
+        const text = await upstreamRes.text();
+        logAiUsageAndCost({
+            capability,
+            provider,
+            requestedModel,
+            requestBody: req.body,
+            responseText: text,
+            isStream: false,
+            headers: upstreamRes.headers
+        });
+        return res.send(text);
+    } catch (error) {
+        logger.error(error);
+        return res.status(HttpCode.INTERNAL_SERVER_ERROR).json({
+            error: { message: "Failed to proxy inference request" }
+        });
+    }
+}
@@ -0,0 +1,293 @@
+import { Request, Response } from "express";
+import { and, eq } from "drizzle-orm";
+import { AiProvider, db, exitNodes, sites, targets } from "@server/db";
+import config from "@server/lib/config";
+import { decrypt } from "@server/lib/crypto";
+import { localCache } from "@server/lib/cache";
+import {
+    AiProviderAuthType,
+    applyAiProviderAuthHeaders,
+    applyAiProviderCustomHeaders,
+    authTypeRequiresApiKey
+} from "@server/lib/aiProviderDefaults";
+import {
+    AI_CAPABILITY_DEFS,
+    type AiCapability
+} from "@server/lib/aiCapabilities";
+import logger from "@server/logger";
+import HttpCode from "@server/types/HttpCode";
+import {
+    applyRequestUserHeaders,
+    type RequestUser
+} from "@server/routers/aiGateway/pipeline";
+
+// Short TTL: long enough to spare the DB on a burst of requests, short
+// enough that target/site changes (added, removed, exit node moved) show up
+// almost immediately without needing explicit cache invalidation.
+const PROVIDER_TARGETS_TTL_SEC = 7;
+
+// Header gerbil reads to know which scheme://host:port (reachable over the
+// WireGuard network) to rewrite an incoming /router/* request to. Must
+// match gerbil's `pangolinDestHeader` constant.
+const PANGOLIN_DEST_HEADER = "p-dest-header";
+
+// Header gerbil reads for the Host header value to send to the destination,
+// when it should differ from PANGOLIN_DEST_HEADER (the target's configured
+// ip rather than the WireGuard routing address). Must match gerbil's
+// `pangolinHostHeader` constant.
+const PANGOLIN_HOST_HEADER = "p-dest-host-header";
+
+const SKIP_HEADERS = new Set([
+    "p-host",
+    "host",
+    "connection",
+    "keep-alive",
+    "proxy-authenticate",
+    "proxy-authorization",
+    "te",
+    "trailers",
+    "transfer-encoding",
+    "upgrade",
+    "content-length",
+    "accept-encoding"
+]);
+
+type ResolvedProviderTarget = {
+    targetId: number;
+    // "<scheme>://<site exitNodeSubnet host>:<internalPort>", passed to
+    // gerbil as the destination to proxy the request to over the WireGuard
+    // tunnel.
+    destination: string;
+    // The target's configured ip, passed to gerbil as the Host header to
+    // send to the destination (which may differ from the WireGuard routing
+    // address above, e.g. for vhost-based targets).
+    hostHeader: string;
+    // The target's site's exit node HTTP API base URL (gerbil's /router/*).
+    gerbilBaseUrl: string;
+};
+
+async function fetchProviderTargets(
+    providerId: number
+): Promise<ResolvedProviderTarget[]> {
+    const rows = await db
+        .select({
+            targetId: targets.targetId,
+            ip: targets.ip,
+            internalPort: targets.internalPort,
+            port: targets.port,
+            method: targets.method,
+            exitNodeSubnet: sites.exitNodeSubnet,
+            reachableAt: exitNodes.reachableAt
+        })
+        .from(targets)
+        .innerJoin(sites, eq(targets.siteId, sites.siteId))
+        .innerJoin(exitNodes, eq(sites.exitNodeId, exitNodes.exitNodeId))
+        .where(
+            and(eq(targets.providerId, providerId), eq(targets.enabled, true))
+        );
+
+    const resolved: ResolvedProviderTarget[] = [];
+    for (const row of rows) {
+        // Sites not yet connected to an exit node (no subnet assigned) or
+        // whose exit node has no known HTTP address can't be routed to.
+        if (!row.exitNodeSubnet || !row.reachableAt) {
+            continue;
+        }
+        const host = row.exitNodeSubnet.split("/")[0];
+        const port = row.internalPort ?? row.port;
+        const scheme = row.method?.toLowerCase() ?? "https";
+        resolved.push({
+            targetId: row.targetId,
+            destination: `${scheme}://${host}:${port}`,
+            hostHeader: row.ip,
+            gerbilBaseUrl: row.reachableAt
+        });
+    }
+
+    return resolved;
+}
+
+async function getProviderTargets(
+    providerId: number
+): Promise<ResolvedProviderTarget[]> {
+    const cacheKey = `aiGateway:providerTargets:${providerId}`;
+    const cached = localCache.get<ResolvedProviderTarget[]>(cacheKey);
+    if (cached !== undefined) {
+        return cached;
+    }
+
+    const resolved = await fetchProviderTargets(providerId);
+    localCache.set(cacheKey, resolved, PROVIDER_TARGETS_TTL_SEC);
+    return resolved;
+}
+
+// Round-robin cursor per provider. Process-local and unpersisted - fine
+// since it only needs to spread load across targets, not guarantee a
+// perfectly even distribution across restarts or multiple server instances.
+const roundRobinCursors = new Map<number, number>();
+
+function pickTarget(
+    providerId: number,
+    providerTargets: ResolvedProviderTarget[]
+): ResolvedProviderTarget {
+    const cursor = roundRobinCursors.get(providerId) ?? 0;
+    roundRobinCursors.set(providerId, cursor + 1);
+    return providerTargets[cursor % providerTargets.length];
+}
+
+function pathFromRequest(req: Request): string {
+    // Query string is preserved - some providers use it to select the
+    // streaming response format (e.g. Gemini's `?alt=sse`), and gerbil's
+    // /router/* forwards it through untouched.
+    const raw = req.originalUrl || req.url || req.path;
+    return raw.startsWith("/") ? raw : `/${raw}`;
+}
+
+/**
+ * Proxies an AI gateway request to one of a "custom" / "target" routing-mode
+ * provider's site targets, via that site's gerbil sidecar. Gerbil's
+ * /router/* endpoint forwards the request (untouched body, same path minus
+ * the /router prefix, and all headers besides PANGOLIN_DEST_HEADER and
+ * PANGOLIN_HOST_HEADER) over the WireGuard tunnel to the destination named
+ * in PANGOLIN_DEST_HEADER, sending PANGOLIN_HOST_HEADER as the Host header.
+ * Always writes a response to `res`, including on failure.
+ */
+export async function proxyAiGatewayToSiteTarget(
+    req: Request,
+    res: Response,
+    provider: AiProvider,
+    requestUser: RequestUser | null,
+    capability: AiCapability
+): Promise<void> {
+    const providerTargets = await getProviderTargets(provider.providerId);
+    if (providerTargets.length === 0) {
+        res.status(HttpCode.INTERNAL_SERVER_ERROR).json({
+            error: {
+                message: "AI provider has no reachable site targets configured"
+            }
+        });
+        return;
+    }
+
+    const target = pickTarget(provider.providerId, providerTargets);
+    const gerbilUrl = `${target.gerbilBaseUrl.replace(/\/+$/, "")}/router${pathFromRequest(req)}`;
+
+    const headers: Record<string, string> = {};
+    for (const [key, value] of Object.entries(req.headers)) {
+        if (SKIP_HEADERS.has(key.toLowerCase()) || value === undefined) {
+            continue;
+        }
+        headers[key] = Array.isArray(value) ? value.join(", ") : value;
+    }
+
+    const authType = provider.authType as AiProviderAuthType;
+    let apiKey: string | null = null;
+    if (authTypeRequiresApiKey(authType)) {
+        if (!provider.apiKey) {
+            res.status(HttpCode.INTERNAL_SERVER_ERROR).json({
+                error: {
+                    message: "AI provider has no API key configured"
+                }
+            });
+            return;
+        }
+        const secret = config.getRawConfig().server.secret!;
+        apiKey = decrypt(provider.apiKey, secret);
+    }
+    applyAiProviderCustomHeaders(
+        headers,
+        provider.headers,
+        config.getRawConfig().server.secret!
+    );
+    applyAiProviderAuthHeaders(headers, authType, apiKey);
+    applyRequestUserHeaders(headers, requestUser);
+
+    headers[PANGOLIN_DEST_HEADER] = target.destination;
+    headers[PANGOLIN_HOST_HEADER] = target.hostHeader;
+
+    const body = JSON.stringify(req.body);
+
+    logger.debug("AI gateway target-routed request", {
+        providerId: provider.providerId,
+        targetId: target.targetId,
+        destination: target.destination,
+        hostHeader: target.hostHeader,
+        url: gerbilUrl,
+        headers,
+        body: req.body
+    });
+
+    // Cancel the request to gerbil (which cascades to gerbil cancelling its
+    // proxied request to the actual site target, since gerbil's reverse
+    // proxy derives the outbound request's context from the inbound one) if
+    // the client goes away before we're done.
+    const abortController = new AbortController();
+    const onClientClose = () => {
+        if (!res.writableEnded) {
+            abortController.abort();
+        }
+    };
+    res.on("close", onClientClose);
+
+    let upstreamRes: globalThis.Response;
+    try {
+        upstreamRes = await fetch(gerbilUrl, {
+            method: "POST",
+            headers,
+            body,
+            signal: abortController.signal
+        });
+    } catch (fetchError) {
+        res.off("close", onClientClose);
+        if (abortController.signal.aborted) {
+            // Client already disconnected; nothing left to respond to.
+            return;
+        }
+        logger.error({
+            message: "AI gateway target proxy request failed",
+            url: gerbilUrl,
+            targetId: target.targetId,
+            error: fetchError,
+            cause:
+                fetchError instanceof Error
+                    ? (fetchError as Error & { cause?: unknown }).cause
+                    : undefined
+        });
+        res.status(HttpCode.BAD_GATEWAY).json({
+            error: { message: "Failed to reach AI provider target" }
+        });
+        return;
+    }
+
+    const contentType = upstreamRes.headers.get("content-type") || "";
+    const isStream = AI_CAPABILITY_DEFS[capability].isStreaming(
+        req,
+        contentType
+    );
+
+    res.status(upstreamRes.status);
+    res.setHeader("Content-Type", contentType || "application/json");
+
+    if (isStream && upstreamRes.body) {
+        res.flushHeaders();
+        const reader = upstreamRes.body.getReader();
+        try {
+            while (!abortController.signal.aborted) {
+                const { done, value } = await reader.read();
+                if (done) break;
+                res.write(value);
+            }
+        } finally {
+            await reader.cancel().catch(() => {});
+            res.off("close", onClientClose);
+        }
+        if (!res.writableEnded) {
+            res.end();
+        }
+        return;
+    }
+
+    res.off("close", onClientClose);
+    const text = await upstreamRes.text();
+    res.send(text);
+}
@@ -0,0 +1,141 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { aiModels, aiProviders, db } from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { and, eq } from "drizzle-orm";
+import type { CreateOrEditAiModelResponse } from "@server/routers/aiProvider/types";
+import { modelListTypeSchema } from "@server/lib/aiInferenceResource";
+
+const paramsSchema = z.strictObject({
+    providerId: z.coerce.number().int().positive()
+});
+
+const bodySchema = z.strictObject({
+    modelKey: z.string().nonempty(),
+    name: z.string().nonempty(),
+    enabled: z.boolean().optional(),
+    listType: modelListTypeSchema.optional().default("allow")
+});
+
+registry.registerPath({
+    method: "put",
+    path: "/ai-provider/{providerId}/model",
+    description: "Create an AI model under a provider.",
+    tags: [OpenAPITags.AiModel],
+    request: {
+        params: paramsSchema,
+        body: {
+            content: {
+                "application/json": {
+                    schema: bodySchema
+                }
+            }
+        }
+    },
+    responses: {
+        201: {
+            description: "Successful response"
+        }
+    }
+});
+
+export async function createAiModel(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const parsedBody = bodySchema.safeParse(req.body);
+        if (!parsedBody.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedBody.error).toString()
+                )
+            );
+        }
+
+        const { providerId } = parsedParams.data;
+        const { modelKey, name, enabled, listType } = parsedBody.data;
+
+        const [provider] =
+            req.aiProvider && req.aiProvider.providerId === providerId
+                ? [req.aiProvider]
+                : await db
+                      .select()
+                      .from(aiProviders)
+                      .where(eq(aiProviders.providerId, providerId))
+                      .limit(1);
+
+        if (!provider) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    `AI provider with ID ${providerId} not found`
+                )
+            );
+        }
+
+        const [existing] = await db
+            .select({ modelId: aiModels.modelId })
+            .from(aiModels)
+            .where(
+                and(
+                    eq(aiModels.providerId, providerId),
+                    eq(aiModels.modelKey, modelKey)
+                )
+            )
+            .limit(1);
+
+        if (existing) {
+            return next(
+                createHttpError(
+                    HttpCode.CONFLICT,
+                    `Model with key ${modelKey} already exists for this provider`
+                )
+            );
+        }
+
+        const now = Date.now();
+        const [model] = await db
+            .insert(aiModels)
+            .values({
+                providerId,
+                modelKey,
+                name,
+                listType,
+                enabled: enabled ?? true,
+                createdAt: now,
+                updatedAt: now
+            })
+            .returning();
+
+        return response<CreateOrEditAiModelResponse>(res, {
+            data: { model },
+            success: true,
+            error: false,
+            message: "AI model created successfully",
+            status: HttpCode.CREATED
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}
@@ -0,0 +1,171 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { aiProviders, db } from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { encrypt } from "@server/lib/crypto";
+import config from "@server/lib/config";
+import {
+    resolveAiProviderCreateFields,
+    resolveCapabilitiesForCreate,
+    serializeAiProviderHeaders
+} from "@server/lib/aiProviderDefaults";
+import type { CreateOrEditAiProviderResponse } from "@server/routers/aiProvider/types";
+import { toPublicAiProvider } from "@server/routers/aiProvider/types";
+import {
+    aiAuthTypeSchema,
+    aiCapabilitiesSchema,
+    aiProviderHeadersSchema,
+    aiProviderTypeSchema,
+    aiRoutingModeSchema,
+    refineProviderUpstreamFields
+} from "@server/routers/aiProvider/validation";
+import { serializeCapabilities } from "@server/lib/aiCapabilities";
+
+const paramsSchema = z.strictObject({
+    orgId: z.string().nonempty()
+});
+
+const bodySchema = z
+    .strictObject({
+        name: z.string().nonempty(),
+        type: aiProviderTypeSchema,
+        upstreamUrl: z.url().optional().nullable(),
+        apiKey: z.string().optional(),
+        authType: aiAuthTypeSchema.optional(),
+        routingMode: aiRoutingModeSchema.optional(),
+        capabilities: aiCapabilitiesSchema.optional(),
+        headers: aiProviderHeadersSchema,
+        skipTlsVerification: z.boolean().optional(),
+        enabled: z.boolean().optional()
+    })
+    .superRefine((data, ctx) => {
+        refineProviderUpstreamFields(data, ctx);
+    });
+
+registry.registerPath({
+    method: "put",
+    path: "/org/{orgId}/ai-provider",
+    description: "Create an AI provider for an organization.",
+    tags: [OpenAPITags.AiProvider],
+    request: {
+        params: paramsSchema,
+        body: {
+            content: {
+                "application/json": {
+                    schema: bodySchema
+                }
+            }
+        }
+    },
+    responses: {
+        201: {
+            description: "Successful response"
+        }
+    }
+});
+
+export async function createAiProvider(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const parsedBody = bodySchema.safeParse(req.body);
+        if (!parsedBody.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedBody.error).toString()
+                )
+            );
+        }
+
+        const { orgId } = parsedParams.data;
+        const {
+            name,
+            type,
+            upstreamUrl,
+            apiKey,
+            authType,
+            routingMode,
+            capabilities,
+            headers,
+            skipTlsVerification,
+            enabled
+        } = parsedBody.data;
+
+        const key = config.getRawConfig().server.secret!;
+        const encryptedApiKey = apiKey ? encrypt(apiKey, key) : null;
+        const apiKeyLastChars = apiKey ? apiKey.slice(-4) : null;
+        const now = Date.now();
+        const resolved = resolveAiProviderCreateFields({
+            type,
+            upstreamUrl,
+            authType,
+            routingMode
+        });
+        const resolvedCapabilities = resolveCapabilitiesForCreate({
+            type,
+            capabilities
+        });
+
+        if (resolvedCapabilities.length === 0) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    "At least one capability is required"
+                )
+            );
+        }
+
+        const [provider] = await db
+            .insert(aiProviders)
+            .values({
+                orgId,
+                name,
+                type,
+                upstreamUrl: resolved.upstreamUrl,
+                apiKey: encryptedApiKey,
+                apiKeyLastChars,
+                authType: resolved.authType,
+                routingMode: resolved.routingMode,
+                capabilities: serializeCapabilities(resolvedCapabilities),
+                headers: serializeAiProviderHeaders(headers, key),
+                skipTlsVerification: skipTlsVerification ?? false,
+                enabled: enabled ?? true,
+                createdAt: now,
+                updatedAt: now
+            })
+            .returning();
+
+        return response<CreateOrEditAiProviderResponse>(res, {
+            data: {
+                provider: toPublicAiProvider(provider, { includeApiKey: true })
+            },
+            success: true,
+            error: false,
+            message: "AI provider created successfully",
+            status: HttpCode.CREATED
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}
@@ -0,0 +1,79 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { aiModels, db } from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { eq } from "drizzle-orm";
+
+const paramsSchema = z.strictObject({
+    modelId: z.coerce.number().int().positive()
+});
+
+registry.registerPath({
+    method: "delete",
+    path: "/ai-model/{modelId}",
+    description: "Delete an AI model.",
+    tags: [OpenAPITags.AiModel],
+    request: {
+        params: paramsSchema
+    },
+    responses: {
+        200: {
+            description: "Successful response"
+        }
+    }
+});
+
+export async function deleteAiModel(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { modelId } = parsedParams.data;
+
+        const [existing] = await db
+            .select({ modelId: aiModels.modelId })
+            .from(aiModels)
+            .where(eq(aiModels.modelId, modelId))
+            .limit(1);
+
+        if (!existing) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    `AI model with ID ${modelId} not found`
+                )
+            );
+        }
+
+        await db.delete(aiModels).where(eq(aiModels.modelId, modelId));
+
+        return response(res, {
+            data: null,
+            success: true,
+            error: false,
+            message: "AI model deleted successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}
@@ -0,0 +1,81 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { aiProviders, db } from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { eq } from "drizzle-orm";
+
+const paramsSchema = z.strictObject({
+    providerId: z.coerce.number().int().positive()
+});
+
+registry.registerPath({
+    method: "delete",
+    path: "/ai-provider/{providerId}",
+    description: "Delete an AI provider.",
+    tags: [OpenAPITags.AiProvider],
+    request: {
+        params: paramsSchema
+    },
+    responses: {
+        200: {
+            description: "Successful response"
+        }
+    }
+});
+
+export async function deleteAiProvider(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { providerId } = parsedParams.data;
+
+        const [existing] = await db
+            .select({ providerId: aiProviders.providerId })
+            .from(aiProviders)
+            .where(eq(aiProviders.providerId, providerId))
+            .limit(1);
+
+        if (!existing) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    `AI provider with ID ${providerId} not found`
+                )
+            );
+        }
+
+        await db
+            .delete(aiProviders)
+            .where(eq(aiProviders.providerId, providerId));
+
+        return response(res, {
+            data: null,
+            success: true,
+            error: false,
+            message: "AI provider deleted successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}
@@ -0,0 +1,81 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { aiModels, db } from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { eq } from "drizzle-orm";
+import type { GetAiModelResponse } from "@server/routers/aiProvider/types";
+
+const paramsSchema = z.strictObject({
+    modelId: z.coerce.number().int().positive()
+});
+
+registry.registerPath({
+    method: "get",
+    path: "/ai-model/{modelId}",
+    description: "Get an AI model by ID.",
+    tags: [OpenAPITags.AiModel],
+    request: {
+        params: paramsSchema
+    },
+    responses: {
+        200: {
+            description: "Successful response"
+        }
+    }
+});
+
+export async function getAiModel(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { modelId } = parsedParams.data;
+
+        const [model] =
+            req.aiModel && req.aiModel.modelId === modelId
+                ? [req.aiModel]
+                : await db
+                      .select()
+                      .from(aiModels)
+                      .where(eq(aiModels.modelId, modelId))
+                      .limit(1);
+
+        if (!model) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    `AI model with ID ${modelId} not found`
+                )
+            );
+        }
+
+        return response<GetAiModelResponse>(res, {
+            data: { model },
+            success: true,
+            error: false,
+            message: "AI model retrieved successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}
@@ -0,0 +1,84 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { aiProviders, db } from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { eq } from "drizzle-orm";
+import type { GetAiProviderResponse } from "@server/routers/aiProvider/types";
+import { toPublicAiProvider } from "@server/routers/aiProvider/types";
+
+const paramsSchema = z.strictObject({
+    providerId: z.coerce.number().int().positive()
+});
+
+registry.registerPath({
+    method: "get",
+    path: "/ai-provider/{providerId}",
+    description: "Get an AI provider by ID.",
+    tags: [OpenAPITags.AiProvider],
+    request: {
+        params: paramsSchema
+    },
+    responses: {
+        200: {
+            description: "Successful response"
+        }
+    }
+});
+
+export async function getAiProvider(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { providerId } = parsedParams.data;
+
+        const [provider] =
+            req.aiProvider && req.aiProvider.providerId === providerId
+                ? [req.aiProvider]
+                : await db
+                      .select()
+                      .from(aiProviders)
+                      .where(eq(aiProviders.providerId, providerId))
+                      .limit(1);
+
+        if (!provider) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    `AI provider with ID ${providerId} not found`
+                )
+            );
+        }
+
+        return response<GetAiProviderResponse>(res, {
+            data: {
+                provider: toPublicAiProvider(provider, { includeApiKey: true })
+            },
+            success: true,
+            error: false,
+            message: "AI provider retrieved successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}
@@ -0,0 +1,11 @@
+export * from "./createAiProvider";
+export * from "./listAiProviders";
+export * from "./getAiProvider";
+export * from "./updateAiProvider";
+export * from "./deleteAiProvider";
+export * from "./createAiModel";
+export * from "./listAiModels";
+export * from "./getAiModel";
+export * from "./updateAiModel";
+export * from "./deleteAiModel";
+export * from "./types";
@@ -0,0 +1,160 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { aiModels, aiProviders, db } from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { and, asc, eq, like, sql } from "drizzle-orm";
+import type { ListAiModelsResponse } from "@server/routers/aiProvider/types";
+
+const paramsSchema = z.strictObject({
+    providerId: z.coerce.number().int().positive()
+});
+
+const listSchema = z.object({
+    pageSize: z.coerce
+        .number<string>()
+        .int()
+        .positive()
+        .optional()
+        .catch(20)
+        .default(20)
+        .openapi({
+            type: "integer",
+            default: 20,
+            description: "Number of items per page"
+        }),
+    page: z.coerce
+        .number<string>()
+        .int()
+        .min(0)
+        .optional()
+        .catch(1)
+        .default(1)
+        .openapi({
+            type: "integer",
+            default: 1,
+            description: "Page number to retrieve"
+        }),
+    query: z.string().optional()
+});
+
+registry.registerPath({
+    method: "get",
+    path: "/ai-provider/{providerId}/models",
+    description: "List AI models for a provider.",
+    tags: [OpenAPITags.AiModel],
+    request: {
+        params: paramsSchema,
+        query: listSchema
+    },
+    responses: {
+        200: {
+            description: "Successful response"
+        }
+    }
+});
+
+export async function listAiModels(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedQuery = listSchema.safeParse(req.query);
+        if (!parsedQuery.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedQuery.error).toString()
+                )
+            );
+        }
+
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { providerId } = parsedParams.data;
+
+        const [provider] =
+            req.aiProvider && req.aiProvider.providerId === providerId
+                ? [req.aiProvider]
+                : await db
+                      .select({ providerId: aiProviders.providerId })
+                      .from(aiProviders)
+                      .where(eq(aiProviders.providerId, providerId))
+                      .limit(1);
+
+        if (!provider) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    `AI provider with ID ${providerId} not found`
+                )
+            );
+        }
+
+        const { pageSize, page, query } = parsedQuery.data;
+        const conditions = [eq(aiModels.providerId, providerId)];
+
+        if (query) {
+            conditions.push(
+                like(
+                    sql`LOWER(${aiModels.name})`,
+                    "%" + query.toLowerCase() + "%"
+                )
+            );
+        }
+
+        const baseQuery = db
+            .select()
+            .from(aiModels)
+            .where(and(...conditions));
+
+        const countQuery = db.$count(
+            db
+                .select()
+                .from(aiModels)
+                .where(and(...conditions))
+                .as("filtered_ai_models")
+        );
+
+        const [totalCount, rows] = await Promise.all([
+            countQuery,
+            baseQuery
+                .limit(pageSize)
+                .offset(pageSize * (page - 1))
+                .orderBy(asc(aiModels.name))
+        ]);
+
+        return response<ListAiModelsResponse>(res, {
+            data: {
+                models: rows,
+                pagination: {
+                    total: totalCount,
+                    pageSize,
+                    page
+                }
+            },
+            success: true,
+            error: false,
+            message: "AI models retrieved successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}
@@ -0,0 +1,152 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { aiProviders, db } from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { and, asc, eq, like, sql } from "drizzle-orm";
+import type { ListAiProvidersResponse } from "@server/routers/aiProvider/types";
+import { toPublicAiProvider } from "@server/routers/aiProvider/types";
+
+const paramsSchema = z.strictObject({
+    orgId: z.string().nonempty()
+});
+
+const listSchema = z.object({
+    pageSize: z.coerce
+        .number<string>()
+        .int()
+        .positive()
+        .optional()
+        .catch(20)
+        .default(20)
+        .openapi({
+            type: "integer",
+            default: 20,
+            description: "Number of items per page"
+        }),
+    page: z.coerce
+        .number<string>()
+        .int()
+        .min(0)
+        .optional()
+        .catch(1)
+        .default(1)
+        .openapi({
+            type: "integer",
+            default: 1,
+            description: "Page number to retrieve"
+        }),
+    query: z.string().optional()
+});
+
+registry.registerPath({
+    method: "get",
+    path: "/org/{orgId}/ai-providers",
+    description: "List AI providers for an organization.",
+    tags: [OpenAPITags.AiProvider],
+    request: {
+        params: paramsSchema,
+        query: listSchema
+    },
+    responses: {
+        200: {
+            description: "Successful response"
+        }
+    }
+});
+
+export async function listAiProviders(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedQuery = listSchema.safeParse(req.query);
+        if (!parsedQuery.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedQuery.error).toString()
+                )
+            );
+        }
+
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { orgId } = parsedParams.data;
+
+        if (req.user && orgId && orgId !== req.userOrgId) {
+            return next(
+                createHttpError(
+                    HttpCode.FORBIDDEN,
+                    "User does not have access to this organization"
+                )
+            );
+        }
+
+        const { pageSize, page, query } = parsedQuery.data;
+        const conditions = [eq(aiProviders.orgId, orgId)];
+
+        if (query) {
+            conditions.push(
+                like(
+                    sql`LOWER(${aiProviders.name})`,
+                    "%" + query.toLowerCase() + "%"
+                )
+            );
+        }
+
+        const baseQuery = db
+            .select()
+            .from(aiProviders)
+            .where(and(...conditions));
+
+        const countQuery = db.$count(
+            db
+                .select()
+                .from(aiProviders)
+                .where(and(...conditions))
+                .as("filtered_ai_providers")
+        );
+
+        const [totalCount, rows] = await Promise.all([
+            countQuery,
+            baseQuery
+                .limit(pageSize)
+                .offset(pageSize * (page - 1))
+                .orderBy(asc(aiProviders.name))
+        ]);
+
+        return response<ListAiProvidersResponse>(res, {
+            data: {
+                providers: rows.map((row) => toPublicAiProvider(row)),
+                pagination: {
+                    total: totalCount,
+                    pageSize,
+                    page
+                }
+            },
+            success: true,
+            error: false,
+            message: "AI providers retrieved successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}
@@ -0,0 +1,86 @@
+import type { AiModel, AiProvider } from "@server/db";
+import type { PaginatedResponse } from "@server/types/Pagination";
+import {
+    parseAiProviderHeaders,
+    type AiProviderAuthType,
+    type AiProviderHeader
+} from "@server/lib/aiProviderDefaults";
+import {
+    parseCapabilities,
+    type AiCapability
+} from "@server/lib/aiCapabilities";
+import { decrypt } from "@server/lib/crypto";
+import config from "@server/lib/config";
+
+export type AiProviderPublic = Omit<
+    AiProvider,
+    "apiKey" | "capabilities" | "headers"
+> & {
+    apiKey?: string | null;
+    capabilities: AiCapability[];
+    headers: AiProviderHeader[] | null;
+    effectiveUpstreamUrl: string | null;
+    effectiveAuthType: AiProviderAuthType;
+};
+
+export type ListAiProvidersResponse = PaginatedResponse<{
+    providers: AiProviderPublic[];
+}>;
+
+export type GetAiProviderResponse = {
+    provider: AiProviderPublic;
+};
+
+export type CreateOrEditAiProviderResponse = {
+    provider: AiProviderPublic;
+};
+
+export type ListAiModelsResponse = PaginatedResponse<{
+    models: AiModel[];
+}>;
+
+export type GetAiModelResponse = {
+    model: AiModel;
+};
+
+export type CreateOrEditAiModelResponse = {
+    model: AiModel;
+};
+
+export function toPublicAiProvider(
+    provider: AiProvider,
+    options?: { includeApiKey?: boolean }
+): AiProviderPublic {
+    const {
+        apiKey: encryptedApiKey,
+        capabilities: rawCapabilities,
+        headers: rawHeaders,
+        ...rest
+    } = provider;
+
+    let apiKey: string | null | undefined;
+    if (options?.includeApiKey) {
+        if (encryptedApiKey) {
+            apiKey = decrypt(
+                encryptedApiKey,
+                config.getRawConfig().server.secret!
+            );
+        } else {
+            apiKey = null;
+        }
+    }
+
+    const parsedHeaders = parseAiProviderHeaders(
+        rawHeaders,
+        config.getRawConfig().server.secret!
+    );
+
+    return {
+        ...rest,
+        ...(options?.includeApiKey ? { apiKey } : {}),
+        capabilities: parseCapabilities(rawCapabilities),
+        headers: parsedHeaders.length > 0 ? parsedHeaders : null,
+        effectiveUpstreamUrl: provider.upstreamUrl,
+        effectiveAuthType: provider.authType as AiProviderAuthType
+    };
+}
@@ -0,0 +1,156 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { aiModels, db } from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { and, eq, ne } from "drizzle-orm";
+import type { CreateOrEditAiModelResponse } from "@server/routers/aiProvider/types";
+import { modelListTypeSchema } from "@server/lib/aiInferenceResource";
+
+const paramsSchema = z.strictObject({
+    modelId: z.coerce.number().int().positive()
+});
+
+const bodySchema = z.strictObject({
+    modelKey: z.string().nonempty().optional(),
+    name: z.string().nonempty().optional(),
+    enabled: z.boolean().optional(),
+    listType: modelListTypeSchema.optional()
+});
+
+registry.registerPath({
+    method: "post",
+    path: "/ai-model/{modelId}",
+    description: "Update an AI model.",
+    tags: [OpenAPITags.AiModel],
+    request: {
+        params: paramsSchema,
+        body: {
+            content: {
+                "application/json": {
+                    schema: bodySchema
+                }
+            }
+        }
+    },
+    responses: {
+        200: {
+            description: "Successful response"
+        }
+    }
+});
+
+export async function updateAiModel(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const parsedBody = bodySchema.safeParse(req.body);
+        if (!parsedBody.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedBody.error).toString()
+                )
+            );
+        }
+
+        const { modelId } = parsedParams.data;
+        const body = parsedBody.data;
+
+        const [existing] =
+            req.aiModel && req.aiModel.modelId === modelId
+                ? [req.aiModel]
+                : await db
+                      .select()
+                      .from(aiModels)
+                      .where(eq(aiModels.modelId, modelId))
+                      .limit(1);
+
+        if (!existing) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    `AI model with ID ${modelId} not found`
+                )
+            );
+        }
+
+        if (
+            body.modelKey !== undefined &&
+            body.modelKey !== existing.modelKey
+        ) {
+            const [conflict] = await db
+                .select({ modelId: aiModels.modelId })
+                .from(aiModels)
+                .where(
+                    and(
+                        eq(aiModels.providerId, existing.providerId),
+                        eq(aiModels.modelKey, body.modelKey),
+                        ne(aiModels.modelId, modelId)
+                    )
+                )
+                .limit(1);
+
+            if (conflict) {
+                return next(
+                    createHttpError(
+                        HttpCode.CONFLICT,
+                        `Model with key ${body.modelKey} already exists for this provider`
+                    )
+                );
+            }
+        }
+
+        const updateData: Partial<typeof aiModels.$inferInsert> = {
+            updatedAt: Date.now()
+        };
+
+        if (body.modelKey !== undefined) {
+            updateData.modelKey = body.modelKey;
+        }
+        if (body.name !== undefined) {
+            updateData.name = body.name;
+        }
+        if (body.enabled !== undefined) {
+            updateData.enabled = body.enabled;
+        }
+        if (body.listType !== undefined) {
+            updateData.listType = body.listType;
+        }
+
+        const [model] = await db
+            .update(aiModels)
+            .set(updateData)
+            .where(eq(aiModels.modelId, modelId))
+            .returning();
+
+        return response<CreateOrEditAiModelResponse>(res, {
+            data: { model },
+            success: true,
+            error: false,
+            message: "AI model updated successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}
@@ -0,0 +1,223 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { aiProviders, db } from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { eq } from "drizzle-orm";
+import { encrypt } from "@server/lib/crypto";
+import config from "@server/lib/config";
+import type { CreateOrEditAiProviderResponse } from "@server/routers/aiProvider/types";
+import { toPublicAiProvider } from "@server/routers/aiProvider/types";
+import {
+    aiAuthTypeSchema,
+    aiCapabilitiesSchema,
+    aiProviderHeadersSchema,
+    aiProviderTypeSchema,
+    aiRoutingModeSchema,
+    refineProviderUpstreamFields
+} from "@server/routers/aiProvider/validation";
+import {
+    serializeAiProviderHeaders,
+    type AiProviderAuthType,
+    type AiProviderRoutingMode,
+    type AiProviderType
+} from "@server/lib/aiProviderDefaults";
+import {
+    parseCapabilities,
+    serializeCapabilities
+} from "@server/lib/aiCapabilities";
+
+const paramsSchema = z.strictObject({
+    providerId: z.coerce.number().int().positive()
+});
+
+const bodySchema = z.strictObject({
+    name: z.string().nonempty().optional(),
+    upstreamUrl: z.url().optional().nullable(),
+    apiKey: z.string().optional(),
+    authType: aiAuthTypeSchema.optional(),
+    routingMode: aiRoutingModeSchema.optional(),
+    capabilities: aiCapabilitiesSchema.optional(),
+    headers: aiProviderHeadersSchema,
+    skipTlsVerification: z.boolean().optional(),
+    enabled: z.boolean().optional()
+});
+
+registry.registerPath({
+    method: "post",
+    path: "/ai-provider/{providerId}",
+    description: "Update an AI provider.",
+    tags: [OpenAPITags.AiProvider],
+    request: {
+        params: paramsSchema,
+        body: {
+            content: {
+                "application/json": {
+                    schema: bodySchema
+                }
+            }
+        }
+    },
+    responses: {
+        200: {
+            description: "Successful response"
+        }
+    }
+});
+
+export async function updateAiProvider(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const parsedBody = bodySchema.safeParse(req.body);
+        if (!parsedBody.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedBody.error).toString()
+                )
+            );
+        }
+
+        const { providerId } = parsedParams.data;
+        const body = parsedBody.data;
+
+        const [existing] =
+            req.aiProvider && req.aiProvider.providerId === providerId
+                ? [req.aiProvider]
+                : await db
+                      .select()
+                      .from(aiProviders)
+                      .where(eq(aiProviders.providerId, providerId))
+                      .limit(1);
+
+        if (!existing) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    `AI provider with ID ${providerId} not found`
+                )
+            );
+        }
+
+        const providerType = existing.type as AiProviderType;
+        const nextRoutingMode: AiProviderRoutingMode =
+            providerType === "custom"
+                ? ((body.routingMode ??
+                      existing.routingMode) as AiProviderRoutingMode)
+                : "url";
+        const nextUpstreamUrl =
+            body.upstreamUrl !== undefined
+                ? body.upstreamUrl
+                : existing.upstreamUrl;
+        const nextAuthType: AiProviderAuthType =
+            body.authType !== undefined
+                ? body.authType
+                : (existing.authType as AiProviderAuthType);
+
+        const nextCapabilities =
+            body.capabilities !== undefined
+                ? body.capabilities
+                : parseCapabilities(existing.capabilities);
+
+        const validation = z
+            .object({
+                type: aiProviderTypeSchema,
+                upstreamUrl: z.string().nullable().optional(),
+                authType: aiAuthTypeSchema,
+                routingMode: aiRoutingModeSchema.optional(),
+                capabilities: aiCapabilitiesSchema.optional()
+            })
+            .superRefine((data, ctx) => refineProviderUpstreamFields(data, ctx))
+            .safeParse({
+                type: providerType,
+                upstreamUrl: nextUpstreamUrl,
+                authType: nextAuthType,
+                routingMode: nextRoutingMode,
+                capabilities: nextCapabilities
+            });
+
+        if (!validation.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(validation.error).toString()
+                )
+            );
+        }
+
+        const updateData: Partial<typeof aiProviders.$inferInsert> = {
+            updatedAt: Date.now(),
+            routingMode: nextRoutingMode
+        };
+
+        if (body.name !== undefined) {
+            updateData.name = body.name;
+        }
+        if (body.skipTlsVerification !== undefined) {
+            updateData.skipTlsVerification = body.skipTlsVerification;
+        }
+        if (body.enabled !== undefined) {
+            updateData.enabled = body.enabled;
+        }
+        if (nextRoutingMode === "target") {
+            updateData.upstreamUrl = null;
+        } else if (body.upstreamUrl !== undefined) {
+            updateData.upstreamUrl = body.upstreamUrl;
+        }
+        if (body.authType !== undefined) {
+            updateData.authType = body.authType;
+        }
+        if (body.capabilities !== undefined) {
+            updateData.capabilities = serializeCapabilities(body.capabilities);
+        }
+
+        if (body.apiKey !== undefined) {
+            const key = config.getRawConfig().server.secret!;
+            updateData.apiKey = encrypt(body.apiKey, key);
+            updateData.apiKeyLastChars = body.apiKey.slice(-4);
+        }
+
+        if (body.headers !== undefined) {
+            const key = config.getRawConfig().server.secret!;
+            updateData.headers = serializeAiProviderHeaders(body.headers, key);
+        }
+
+        const [provider] = await db
+            .update(aiProviders)
+            .set(updateData)
+            .where(eq(aiProviders.providerId, providerId))
+            .returning();
+
+        return response<CreateOrEditAiProviderResponse>(res, {
+            data: {
+                provider: toPublicAiProvider(provider, { includeApiKey: true })
+            },
+            success: true,
+            error: false,
+            message: "AI provider updated successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}
@@ -0,0 +1,126 @@
+import { z } from "zod";
+import {
+    AI_PROVIDER_AUTH_TYPES,
+    providerRequiresUpstreamUrl,
+    type AiProviderAuthType,
+    type AiProviderRoutingMode,
+    type AiProviderType
+} from "@server/lib/aiProviderDefaults";
+import { AI_CAPABILITIES } from "@server/lib/aiCapabilities";
+
+export const aiProviderTypeSchema = z.enum([
+    "openai",
+    "anthropic",
+    "googleGemini",
+    "vertexAi",
+    "bedrock",
+    "microsoftFoundry",
+    "openRouter",
+    "vercelAiGateway",
+    "custom"
+]);
+
+export const aiAuthTypeSchema = z.enum(AI_PROVIDER_AUTH_TYPES);
+
+export const aiRoutingModeSchema = z.enum(["url", "target"]);
+
+export const aiCapabilitySchema = z.enum(AI_CAPABILITIES);
+
+export const aiCapabilitiesSchema = z.array(aiCapabilitySchema);
+
+const validHeaderName = /^[a-zA-Z0-9!#$%&'*+\-.^_`|~]+$/;
+const validHeaderValue = /^[\t\x20-\x7E]*$/;
+const templatePattern = /\{\{[^}]+\}\}/;
+
+export const aiProviderHeadersSchema = z
+    .array(z.strictObject({ name: z.string(), value: z.string() }))
+    .nullable()
+    .optional()
+    .superRefine((headers, ctx) => {
+        if (!headers) {
+            return;
+        }
+        for (const [index, header] of headers.entries()) {
+            if (!validHeaderName.test(header.name)) {
+                ctx.addIssue({
+                    code: "custom",
+                    message:
+                        "Header names may only contain valid HTTP token characters (letters, digits, and !#$%&'*+-.^_`|~).",
+                    path: [index, "name"]
+                });
+            }
+            if (!validHeaderValue.test(header.value)) {
+                ctx.addIssue({
+                    code: "custom",
+                    message:
+                        "Header values may only contain printable ASCII characters and horizontal whitespace.",
+                    path: [index, "value"]
+                });
+            }
+            if (
+                templatePattern.test(header.name) ||
+                templatePattern.test(header.value)
+            ) {
+                ctx.addIssue({
+                    code: "custom",
+                    message:
+                        "Header names and values must not contain template expressions such as {{value}}.",
+                    path: [index]
+                });
+            }
+        }
+    });
+
+export function refineProviderUpstreamFields(
+    data: {
+        type: AiProviderType;
+        upstreamUrl?: string | null;
+        authType?: AiProviderAuthType | null;
+        routingMode?: AiProviderRoutingMode | null;
+        capabilities?: z.infer<typeof aiCapabilitiesSchema> | null;
+    },
+    ctx: z.RefinementCtx
+) {
+    const routingMode = data.routingMode ?? "url";
+
+    if (data.type !== "custom" && routingMode === "target") {
+        ctx.addIssue({
+            code: "custom",
+            message: "routingMode target is only allowed for custom providers",
+            path: ["routingMode"]
+        });
+    }
+
+    if (
+        providerRequiresUpstreamUrl(data.type, routingMode) &&
+        !data.upstreamUrl
+    ) {
+        ctx.addIssue({
+            code: "custom",
+            message: `upstreamUrl is required for ${data.type} providers`,
+            path: ["upstreamUrl"]
+        });
+    }
+
+    if (data.type === "custom") {
+        const caps = data.capabilities;
+        if (!caps || caps.length === 0) {
+            ctx.addIssue({
+                code: "custom",
+                message:
+                    "At least one capability is required for custom providers",
+                path: ["capabilities"]
+            });
+        }
+    } else if (
+        data.capabilities !== undefined &&
+        data.capabilities !== null &&
+        data.capabilities.length === 0
+    ) {
+        ctx.addIssue({
+            code: "custom",
+            message: "At least one capability is required",
+            path: ["capabilities"]
+        });
+    }
+}
@@ -222,7 +222,9 @@ export async function verifyResourceSession(
        }

        const { blockAccess, mode } = resource;
-        const dontStripSession = ["ssh", "rdp", "vnc"].includes(mode);
+        const dontStripSession = ["ssh", "rdp", "vnc", "inference"].includes(
+            mode
+        );

        if (blockAccess) {
            logger.debug("Resource blocked", host);
@@ -255,20 +255,6 @@ export async function createClient(

        let newClient: Client | null = null;
        await db.transaction(async (trx) => {
-            // TODO: more intelligent way to pick the exit node
-            const exitNodesList = await listExitNodes(orgId);
-            const randomExitNode =
-                exitNodesList[Math.floor(Math.random() * exitNodesList.length)];
-
-            if (!randomExitNode) {
-                return next(
-                    createHttpError(
-                        HttpCode.NOT_FOUND,
-                        `No exit nodes available. ${build == "saas" ? "Please contact support." : "You need to install gerbil to use the clients."}`
-                    )
-                );
-            }
-
            const [adminRole] = await trx
                .select()
                .from(roles)
@@ -287,7 +273,6 @@ export async function createClient(
                .insert(clients)
                .values({
                    niceId,
-                    exitNodeId: randomExitNode.exitNodeId,
                    orgId,
                    name,
                    subnet: updatedSubnet,
@@ -222,11 +222,6 @@ export async function createUserClient(

        let newClient: Client | null = null;
        await db.transaction(async (trx) => {
-            // TODO: more intelligent way to pick the exit node
-            const exitNodesList = await listExitNodes(orgId);
-            const randomExitNode =
-                exitNodesList[Math.floor(Math.random() * exitNodesList.length)];
-
            const [adminRole] = await trx
                .select()
                .from(roles)
@@ -244,7 +239,6 @@ export async function createUserClient(
            [newClient] = await trx
                .insert(clients)
                .values({
-                    exitNodeId: randomExitNode.exitNodeId,
                    orgId,
                    niceId,
                    name,
@@ -45,7 +45,9 @@ import {
    verifySiteResourceAccess,
    verifyOlmAccess,
    verifyLimits,
-    verifyResourcePolicyAccess
+    verifyResourcePolicyAccess,
+    verifyAiProviderAccess,
+    verifyAiModelAccess
 } from "@server/middlewares";
 import { ActionsEnum } from "@server/auth/actions";
 import rateLimit, { ipKeyGenerator } from "express-rate-limit";
@@ -55,6 +57,7 @@ import { createStore } from "#dynamic/lib/rateLimitStore";
 import { logActionAudit } from "#dynamic/middlewares";
 import { checkRoundTripMessage } from "./ws";
 import * as labels from "@server/routers/labels";
+import * as aiProvider from "@server/routers/aiProvider";

 // Root routes
 export const unauthenticated = Router();
@@ -404,6 +407,20 @@ authenticated.get(
    siteResource.listSiteResourceClients
 );

+authenticated.get(
+    "/site-resource/:siteResourceId/ai-models",
+    verifySiteResourceAccess,
+    verifyUserHasAction(ActionsEnum.listResourceAiModels),
+    siteResource.listSiteResourceAiModels
+);
+
+authenticated.get(
+    "/site-resource/:siteResourceId/ai-providers",
+    verifySiteResourceAccess,
+    verifyUserHasAction(ActionsEnum.listResourceAiModels),
+    siteResource.listSiteResourceAiProviders
+);
+
 authenticated.post(
    "/site-resource/:siteResourceId/roles",
    verifySiteResourceAccess,
@@ -414,6 +431,54 @@ authenticated.post(
    siteResource.setSiteResourceRoles
 );

+authenticated.post(
+    "/site-resource/:siteResourceId/ai-models",
+    verifySiteResourceAccess,
+    verifyUserHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    siteResource.setSiteResourceAiModels
+);
+
+authenticated.post(
+    "/site-resource/:siteResourceId/ai-models/add",
+    verifySiteResourceAccess,
+    verifyUserHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    siteResource.addAiModelToSiteResource
+);
+
+authenticated.post(
+    "/site-resource/:siteResourceId/ai-models/remove",
+    verifySiteResourceAccess,
+    verifyUserHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    siteResource.removeAiModelFromSiteResource
+);
+
+authenticated.post(
+    "/site-resource/:siteResourceId/ai-providers",
+    verifySiteResourceAccess,
+    verifyUserHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    siteResource.setSiteResourceAiProviders
+);
+
+authenticated.post(
+    "/site-resource/:siteResourceId/ai-providers/add",
+    verifySiteResourceAccess,
+    verifyUserHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    siteResource.addAiProviderToSiteResource
+);
+
+authenticated.post(
+    "/site-resource/:siteResourceId/ai-providers/remove",
+    verifySiteResourceAccess,
+    verifyUserHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    siteResource.removeAiProviderFromSiteResource
+);
+
 authenticated.post(
    "/site-resource/:siteResourceId/users",
    verifySiteResourceAccess,
@@ -648,6 +713,20 @@ authenticated.get(
    resource.listResourceUsers
 );

+authenticated.get(
+    "/resource/:resourceId/ai-models",
+    verifyResourceAccess,
+    verifyUserHasAction(ActionsEnum.listResourceAiModels),
+    resource.listResourceAiModels
+);
+
+authenticated.get(
+    "/resource/:resourceId/ai-providers",
+    verifyResourceAccess,
+    verifyUserHasAction(ActionsEnum.listResourceAiModels),
+    resource.listResourceAiProviders
+);
+
 authenticated.get(
    "/resource/:resourceId",
    verifyResourceAccess,
@@ -851,6 +930,54 @@ authenticated.post(
    resource.setResourceUsers
 );

+authenticated.post(
+    "/resource/:resourceId/ai-models",
+    verifyResourceAccess,
+    verifyUserHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    resource.setResourceAiModels
+);
+
+authenticated.post(
+    "/resource/:resourceId/ai-models/add",
+    verifyResourceAccess,
+    verifyUserHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    resource.addAiModelToResource
+);
+
+authenticated.post(
+    "/resource/:resourceId/ai-models/remove",
+    verifyResourceAccess,
+    verifyUserHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    resource.removeAiModelFromResource
+);
+
+authenticated.post(
+    "/resource/:resourceId/ai-providers",
+    verifyResourceAccess,
+    verifyUserHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    resource.setResourceAiProviders
+);
+
+authenticated.post(
+    "/resource/:resourceId/ai-providers/add",
+    verifyResourceAccess,
+    verifyUserHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    resource.addAiProviderToResource
+);
+
+authenticated.post(
+    "/resource/:resourceId/ai-providers/remove",
+    verifyResourceAccess,
+    verifyUserHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    resource.removeAiProviderFromResource
+);
+
 authenticated.put(
    "/resource-policy/:resourcePolicyId/access-control",
    verifyResourcePolicyAccess,
@@ -1366,6 +1493,99 @@ authenticated.get(

 authenticated.get("/ws/round-trip-message/:messageId", checkRoundTripMessage);

+authenticated.put(
+    "/org/:orgId/ai-provider",
+    verifyOrgAccess,
+    verifyUserHasAction(ActionsEnum.createAiProvider),
+    logActionAudit(ActionsEnum.createAiProvider),
+    aiProvider.createAiProvider
+);
+
+authenticated.get(
+    "/org/:orgId/ai-providers",
+    verifyOrgAccess,
+    verifyUserHasAction(ActionsEnum.listAiProviders),
+    aiProvider.listAiProviders
+);
+
+authenticated.get(
+    "/ai-provider/:providerId",
+    verifyAiProviderAccess,
+    verifyUserHasAction(ActionsEnum.getAiProvider),
+    aiProvider.getAiProvider
+);
+
+authenticated.put(
+    "/ai-provider/:providerId/target",
+    verifyAiProviderAccess,
+    verifySiteAccess,
+    verifyLimits,
+    verifyUserHasAction(ActionsEnum.createTarget),
+    logActionAudit(ActionsEnum.createTarget),
+    target.createTarget
+);
+
+authenticated.get(
+    "/ai-provider/:providerId/targets",
+    verifyAiProviderAccess,
+    verifyUserHasAction(ActionsEnum.listTargets),
+    target.listTargets
+);
+
+authenticated.post(
+    "/ai-provider/:providerId",
+    verifyAiProviderAccess,
+    verifyUserHasAction(ActionsEnum.updateAiProvider),
+    logActionAudit(ActionsEnum.updateAiProvider),
+    aiProvider.updateAiProvider
+);
+
+authenticated.delete(
+    "/ai-provider/:providerId",
+    verifyAiProviderAccess,
+    verifyUserHasAction(ActionsEnum.deleteAiProvider),
+    logActionAudit(ActionsEnum.deleteAiProvider),
+    aiProvider.deleteAiProvider
+);
+
+authenticated.put(
+    "/ai-provider/:providerId/model",
+    verifyAiProviderAccess,
+    verifyUserHasAction(ActionsEnum.createAiModel),
+    logActionAudit(ActionsEnum.createAiModel),
+    aiProvider.createAiModel
+);
+
+authenticated.get(
+    "/ai-provider/:providerId/models",
+    verifyAiProviderAccess,
+    verifyUserHasAction(ActionsEnum.listAiModels),
+    aiProvider.listAiModels
+);
+
+authenticated.get(
+    "/ai-model/:modelId",
+    verifyAiModelAccess,
+    verifyUserHasAction(ActionsEnum.getAiModel),
+    aiProvider.getAiModel
+);
+
+authenticated.post(
+    "/ai-model/:modelId",
+    verifyAiModelAccess,
+    verifyUserHasAction(ActionsEnum.updateAiModel),
+    logActionAudit(ActionsEnum.updateAiModel),
+    aiProvider.updateAiModel
+);
+
+authenticated.delete(
+    "/ai-model/:modelId",
+    verifyAiModelAccess,
+    verifyUserHasAction(ActionsEnum.deleteAiModel),
+    logActionAudit(ActionsEnum.deleteAiModel),
+    aiProvider.deleteAiModel
+);
+
 authenticated.get(
    "/org/:orgId/labels",
    verifyOrgAccess,
@@ -100,7 +100,7 @@ export async function generateRelayMappings(exitNode: ExitNode) {
    // Filter to sites with the required fields up front so the rest of the
    // function can safely treat endpoint/subnet/listenPort as defined.
    const validSites = sitesRes.filter(
-        (s) => s.endpoint && s.subnet && s.listenPort
+        (s) => s.endpoint && s.exitNodeSubnet && s.listenPort
    );

    if (validSites.length === 0) {
@@ -136,7 +136,7 @@ export async function generateRelayMappings(exitNode: ExitNode) {
        if (
            peer.orgId == null ||
            !peer.endpoint ||
-            !peer.subnet ||
+            !peer.exitNodeSubnet ||
            !peer.listenPort
        ) {
            continue;
@@ -183,7 +183,7 @@ export async function generateRelayMappings(exitNode: ExitNode) {
    // Process each site using the pre-fetched data.
    for (const site of validSites) {
        const siteDestination: PeerDestination = {
-            destinationIP: site.subnet!.split("/")[0],
+            destinationIP: site.exitNodeSubnet!.split("/")[0],
            destinationPort: site.listenPort! || 1 // this satisfies gerbil for now but should be reevaluated
        };

@@ -207,7 +207,7 @@ export async function generateRelayMappings(exitNode: ExitNode) {
                        continue;
                    }
                    addDestination(site.endpoint!, {
-                        destinationIP: peer.subnet!.split("/")[0],
+                        destinationIP: peer.exitNodeSubnet!.split("/")[0],
                        destinationPort: peer.listenPort! || 1 // this satisfies gerbil for now but should be reevaluated
                    });
                }
@@ -1,6 +1,6 @@
 import { Request, Response, NextFunction } from "express";
 import { z } from "zod";
-import { sites, exitNodes, ExitNode } from "@server/db";
+import { sites, exitNodes, ExitNode, clients } from "@server/db";
 import { db } from "@server/db";
 import { eq, isNotNull, and } from "drizzle-orm";
 import HttpCode from "@server/types/HttpCode";
@@ -89,11 +89,27 @@ export async function generateGerbilConfig(exitNode: ExitNode) {
            and(
                eq(sites.exitNodeId, exitNode.exitNodeId),
                isNotNull(sites.pubKey),
-                isNotNull(sites.subnet)
+                isNotNull(sites.exitNodeSubnet)
            )
        );

-    const peers = await Promise.all(
+    const clientsRes = await db
+        .select()
+        .from(clients)
+        .where(
+            and(
+                eq(clients.exitNodeId, exitNode.exitNodeId),
+                isNotNull(clients.pubKey),
+                isNotNull(clients.exitNodeSubnet)
+            )
+        );
+
+    let peers: {
+        publicKey: string | null;
+        allowedIps: string[];
+    }[] = [];
+
+    const sitePeers = await Promise.all(
        sitesRes.map(async (site) => {
            if (site.type === "wireguard") {
                return {
@@ -103,7 +119,7 @@ export async function generateGerbilConfig(exitNode: ExitNode) {
            } else if (site.type === "newt") {
                return {
                    publicKey: site.pubKey,
-                    allowedIps: [site.subnet!]
+                    allowedIps: [site.exitNodeSubnet!]
                };
            }
            return {
@@ -113,6 +129,15 @@ export async function generateGerbilConfig(exitNode: ExitNode) {
        })
    );

+    const clientPeers = clientsRes.map((client) => {
+        return {
+            publicKey: client.pubKey,
+            allowedIps: [client.exitNodeSubnet!]
+        };
+    });
+
+    peers = [...sitePeers, ...clientPeers];
+
    const configResponse: GetConfigResponse = {
        listenPort: exitNode.listenPort || 51820,
        ipAddress: exitNode.address,
@@ -188,7 +188,7 @@ export async function updateAndGenerateEndpointDestinations(
            .select({
                siteId: sites.siteId,
                newtId: newts.newtId,
-                subnet: sites.subnet,
+                subnet: sites.exitNodeSubnet,
                listenPort: sites.listenPort,
                publicKey: sites.publicKey,
                endpoint: clientSitesAssociationsCache.endpoint,
@@ -13,6 +13,7 @@ import * as apiKeys from "./apiKeys";
 import * as idp from "./idp";
 import * as logs from "./auditLogs";
 import * as siteResource from "./siteResource";
+import * as aiProvider from "./aiProvider";
 import {
    verifyApiKey,
    verifyApiKeyOrgAccess,
@@ -31,6 +32,8 @@ import {
    verifyLimits,
    verifyApiKeyDomainAccess,
    verifyApiKeyResourcePolicyAccess,
+    verifyApiKeyAiProviderAccess,
+    verifyApiKeyAiModelAccess,
    verifyUserHasAction
 } from "@server/middlewares";
 import HttpCode from "@server/types/HttpCode";
@@ -243,6 +246,26 @@ authenticated.get(
    siteResource.listSiteResourceClients
 );

+authenticated.get(
+    [
+        "/site-resource/:siteResourceId/ai-models",
+        "/private-resource/:siteResourceId/ai-models"
+    ],
+    verifyApiKeySiteResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.listResourceAiModels),
+    siteResource.listSiteResourceAiModels
+);
+
+authenticated.get(
+    [
+        "/site-resource/:siteResourceId/ai-providers",
+        "/private-resource/:siteResourceId/ai-providers"
+    ],
+    verifyApiKeySiteResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.listResourceAiModels),
+    siteResource.listSiteResourceAiProviders
+);
+
 authenticated.post(
    [
        "/site-resource/:siteResourceId/roles",
@@ -295,6 +318,72 @@ authenticated.post(
    siteResource.removeRoleFromSiteResource
 );

+authenticated.post(
+    [
+        "/site-resource/:siteResourceId/ai-models",
+        "/private-resource/:siteResourceId/ai-models"
+    ],
+    verifyApiKeySiteResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    siteResource.setSiteResourceAiModels
+);
+
+authenticated.post(
+    [
+        "/site-resource/:siteResourceId/ai-models/add",
+        "/private-resource/:siteResourceId/ai-models/add"
+    ],
+    verifyApiKeySiteResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    siteResource.addAiModelToSiteResource
+);
+
+authenticated.post(
+    [
+        "/site-resource/:siteResourceId/ai-models/remove",
+        "/private-resource/:siteResourceId/ai-models/remove"
+    ],
+    verifyApiKeySiteResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    siteResource.removeAiModelFromSiteResource
+);
+
+authenticated.post(
+    [
+        "/site-resource/:siteResourceId/ai-providers",
+        "/private-resource/:siteResourceId/ai-providers"
+    ],
+    verifyApiKeySiteResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    siteResource.setSiteResourceAiProviders
+);
+
+authenticated.post(
+    [
+        "/site-resource/:siteResourceId/ai-providers/add",
+        "/private-resource/:siteResourceId/ai-providers/add"
+    ],
+    verifyApiKeySiteResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    siteResource.addAiProviderToSiteResource
+);
+
+authenticated.post(
+    [
+        "/site-resource/:siteResourceId/ai-providers/remove",
+        "/private-resource/:siteResourceId/ai-providers/remove"
+    ],
+    verifyApiKeySiteResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    siteResource.removeAiProviderFromSiteResource
+);
+
 authenticated.post(
    [
        "/site-resource/:siteResourceId/users/add",
@@ -507,6 +596,26 @@ authenticated.get(
    resource.listResourceUsers
 );

+authenticated.get(
+    [
+        "/resource/:resourceId/ai-models",
+        "/public-resource/:resourceId/ai-models"
+    ],
+    verifyApiKeyResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.listResourceAiModels),
+    resource.listResourceAiModels
+);
+
+authenticated.get(
+    [
+        "/resource/:resourceId/ai-providers",
+        "/public-resource/:resourceId/ai-providers"
+    ],
+    verifyApiKeyResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.listResourceAiModels),
+    resource.listResourceAiProviders
+);
+
 authenticated.get(
    ["/resource/:resourceId", "/public-resource/:resourceId"],
    verifyApiKeyResourceAccess,
@@ -708,6 +817,28 @@ authenticated.post(
    resource.setResourceRoles
 );

+authenticated.post(
+    [
+        "/resource/:resourceId/ai-models",
+        "/public-resource/:resourceId/ai-models"
+    ],
+    verifyApiKeyResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    resource.setResourceAiModels
+);
+
+authenticated.post(
+    [
+        "/resource/:resourceId/ai-providers",
+        "/public-resource/:resourceId/ai-providers"
+    ],
+    verifyApiKeyResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    resource.setResourceAiProviders
+);
+
 authenticated.post(
    ["/resource/:resourceId/users", "/public-resource/:resourceId/users"],
    verifyApiKeyResourceAccess,
@@ -900,6 +1031,50 @@ authenticated.post(
    resource.removeRoleFromResource
 );

+authenticated.post(
+    [
+        "/resource/:resourceId/ai-models/add",
+        "/public-resource/:resourceId/ai-models/add"
+    ],
+    verifyApiKeyResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    resource.addAiModelToResource
+);
+
+authenticated.post(
+    [
+        "/resource/:resourceId/ai-models/remove",
+        "/public-resource/:resourceId/ai-models/remove"
+    ],
+    verifyApiKeyResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    resource.removeAiModelFromResource
+);
+
+authenticated.post(
+    [
+        "/resource/:resourceId/ai-providers/add",
+        "/public-resource/:resourceId/ai-providers/add"
+    ],
+    verifyApiKeyResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    resource.addAiProviderToResource
+);
+
+authenticated.post(
+    [
+        "/resource/:resourceId/ai-providers/remove",
+        "/public-resource/:resourceId/ai-providers/remove"
+    ],
+    verifyApiKeyResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    resource.removeAiProviderFromResource
+);
+
 authenticated.post(
    [
        "/resource/:resourceId/users/add",
@@ -1366,3 +1541,95 @@ authenticated.get(
    verifyApiKeyHasAction(ActionsEnum.listResources),
    resource.listAllResourceNames
 );
+
+authenticated.put(
+    "/org/:orgId/ai-provider",
+    verifyApiKeyOrgAccess,
+    verifyApiKeyHasAction(ActionsEnum.createAiProvider),
+    logActionAudit(ActionsEnum.createAiProvider),
+    aiProvider.createAiProvider
+);
+
+authenticated.get(
+    "/org/:orgId/ai-providers",
+    verifyApiKeyOrgAccess,
+    verifyApiKeyHasAction(ActionsEnum.listAiProviders),
+    aiProvider.listAiProviders
+);
+
+authenticated.get(
+    "/ai-provider/:providerId",
+    verifyApiKeyAiProviderAccess,
+    verifyApiKeyHasAction(ActionsEnum.getAiProvider),
+    aiProvider.getAiProvider
+);
+
+authenticated.put(
+    "/ai-provider/:providerId/target",
+    verifyApiKeyAiProviderAccess,
+    verifyLimits,
+    verifyApiKeyHasAction(ActionsEnum.createTarget),
+    logActionAudit(ActionsEnum.createTarget),
+    target.createTarget
+);
+
+authenticated.get(
+    "/ai-provider/:providerId/targets",
+    verifyApiKeyAiProviderAccess,
+    verifyApiKeyHasAction(ActionsEnum.listTargets),
+    target.listTargets
+);
+
+authenticated.post(
+    "/ai-provider/:providerId",
+    verifyApiKeyAiProviderAccess,
+    verifyApiKeyHasAction(ActionsEnum.updateAiProvider),
+    logActionAudit(ActionsEnum.updateAiProvider),
+    aiProvider.updateAiProvider
+);
+
+authenticated.delete(
+    "/ai-provider/:providerId",
+    verifyApiKeyAiProviderAccess,
+    verifyApiKeyHasAction(ActionsEnum.deleteAiProvider),
+    logActionAudit(ActionsEnum.deleteAiProvider),
+    aiProvider.deleteAiProvider
+);
+
+authenticated.put(
+    "/ai-provider/:providerId/model",
+    verifyApiKeyAiProviderAccess,
+    verifyApiKeyHasAction(ActionsEnum.createAiModel),
+    logActionAudit(ActionsEnum.createAiModel),
+    aiProvider.createAiModel
+);
+
+authenticated.get(
+    "/ai-provider/:providerId/models",
+    verifyApiKeyAiProviderAccess,
+    verifyApiKeyHasAction(ActionsEnum.listAiModels),
+    aiProvider.listAiModels
+);
+
+authenticated.get(
+    "/ai-model/:modelId",
+    verifyApiKeyAiModelAccess,
+    verifyApiKeyHasAction(ActionsEnum.getAiModel),
+    aiProvider.getAiModel
+);
+
+authenticated.post(
+    "/ai-model/:modelId",
+    verifyApiKeyAiModelAccess,
+    verifyApiKeyHasAction(ActionsEnum.updateAiModel),
+    logActionAudit(ActionsEnum.updateAiModel),
+    aiProvider.updateAiModel
+);
+
+authenticated.delete(
+    "/ai-model/:modelId",
+    verifyApiKeyAiModelAccess,
+    verifyApiKeyHasAction(ActionsEnum.deleteAiModel),
+    logActionAudit(ActionsEnum.deleteAiModel),
+    aiProvider.deleteAiModel
+);
@@ -63,3 +63,4 @@ internalRouter.use("/badger", badgerRouter);
 badgerRouter.post("/verify-session", badger.verifyResourceSession);

 badgerRouter.post("/exchange-session", badger.exchangeSession);
+
@@ -1,7 +1,8 @@
+import { SiteResource } from "@server/db";
 import { formatEndpoint, parseEndpoint } from "@server/lib/ip";

 export type SiteResourceDestinationInput = {
-    mode: "host" | "cidr" | "http" | "ssh";
+    mode: SiteResource["mode"];
    destination: string | null;
    destinationPort: number | null;
    scheme: "http" | "https" | null;
@@ -15,7 +15,7 @@ import {
 } from "@server/db";
 import logger from "@server/logger";
 import { initPeerAddHandshake, updatePeer } from "../olm/peers";
-import { eq, and, inArray } from "drizzle-orm";
+import { eq, and, inArray, or, isNotNull, sql } from "drizzle-orm";
 import config from "@server/lib/config";
 import { decrypt } from "@server/lib/crypto";
 import {
@@ -211,7 +211,8 @@ export async function buildClientConfigurationForNewtClient(
    // call rather than letting each resource fetch its own — with thousands
    // of resources this avoids a concurrent DB/cache stampede for what is
    // often the very same (e.g. wildcard) certificate.
-    const certByDomain = await batchFetchCertsForSiteResources(allSiteResources);
+    const certByDomain =
+        await batchFetchCertsForSiteResources(allSiteResources);

    const resourceTargetsArr = await Promise.all(
        allSiteResources.map((resource) =>
@@ -240,7 +241,7 @@ export async function buildTargetConfigurationForNewtClient(
    version?: string | null,
    remoteExitNodeId?: string
 ) {
-    // Get all enabled targets with their resource mode information
+    // Get enabled HTTP/TCP/UDP targets for resources and AI providers
    const allTargets = await db
        .select({
            resourceId: targets.resourceId,
@@ -250,15 +251,18 @@ export async function buildTargetConfigurationForNewtClient(
            port: targets.port,
            internalPort: targets.internalPort,
            enabled: targets.enabled,
-            mode: resources.mode
+            mode: sql<string>`COALESCE(${resources.mode}, ${targets.mode})`.mapWith(
+                String
+            )
        })
        .from(targets)
-        .innerJoin(resources, eq(targets.resourceId, resources.resourceId))
+        .leftJoin(resources, eq(targets.resourceId, resources.resourceId))
        .where(
            and(
                eq(targets.siteId, siteId),
                eq(targets.enabled, true),
-                inArray(targets.mode, ["http", "udp", "tcp"])
+                inArray(targets.mode, ["http", "udp", "tcp"]),
+                or(isNotNull(targets.resourceId), isNotNull(targets.providerId))
            )
        );

@@ -2,14 +2,17 @@ import { db, sites } from "@server/db";
 import { MessageHandler } from "@server/routers/ws";
 import { exitNodes, Newt } from "@server/db";
 import logger from "@server/logger";
-import { ne, eq, or, and, count } from "drizzle-orm";
+import { eq } from "drizzle-orm";
 import { listExitNodes } from "#dynamic/lib/exitNodes";
+import { calculateExitNodeWeight } from "@server/lib/exitNodes";

-export const handleNewtPingRequestMessage: MessageHandler = async (context) => {
+export const handleNewtExitNodesRequestMessage: MessageHandler = async (
+    context
+) => {
    const { message, client, sendToClient } = context;
    const newt = client as Newt;

-    logger.info("Handling ping request newt message!");
+    logger.info("Handling exit nodes request newt message!");

    if (!newt) {
        logger.warn("Newt not found");
@@ -54,32 +57,13 @@ export const handleNewtPingRequestMessage: MessageHandler = async (context) => {

    const exitNodesPayload = await Promise.all(
        exitNodesList.map(async (node) => {
-            // (MAX_CONNECTIONS - current_connections) / MAX_CONNECTIONS)
-            // higher = more desirable
-            // like saying, this node has x% of its capacity left
+            const weight = await calculateExitNodeWeight(
+                node.exitNodeId,
+                node.maxConnections
+            );

-            let weight = 1;
-            const maxConnections = node.maxConnections;
-            if (maxConnections !== null && maxConnections !== undefined) {
-                const [currentConnections] = await db
-                    .select({
-                        count: count()
-                    })
-                    .from(sites)
-                    .where(
-                        and(
-                            eq(sites.exitNodeId, node.exitNodeId),
-                            eq(sites.online, true)
-                        )
-                    );
-
-                if (currentConnections.count >= maxConnections) {
-                    return null;
-                }
-
-                weight =
-                    (maxConnections - currentConnections.count) /
-                    maxConnections;
+            if (weight === null) {
+                return null;
            }

            return {
@@ -95,16 +95,16 @@ export const handleNewtGetConfigMessage: MessageHandler = async (context) => {
            .limit(1);
        if (
            exitNode.reachableAt &&
-            existingSite.subnet &&
+            existingSite.exitNodeSubnet &&
            existingSite.listenPort
        ) {
            const payload = {
                oldDestination: {
-                    destinationIP: existingSite.subnet?.split("/")[0],
+                    destinationIP: existingSite.exitNodeSubnet?.split("/")[0],
                    destinationPort: existingSite.listenPort || 1 // this satisfies gerbil for now but should be reevaluated
                },
                newDestination: {
-                    destinationIP: site.subnet?.split("/")[0],
+                    destinationIP: site.exitNodeSubnet?.split("/")[0],
                    destinationPort: site.listenPort || 1 // this satisfies gerbil for now but should be reevaluated
                }
            };
@@ -132,7 +132,10 @@ export const handleNewtGetConfigMessage: MessageHandler = async (context) => {
        ({ targets: dedupedTargets, certs } = dedupeCertsForTargets(targets));
    }

-    const targetsToSend = await convertTargetsIfNecessary(newt.newtId, dedupedTargets); // for backward compatibility with old newt versions that don't support the new target format
+    const targetsToSend = await convertTargetsIfNecessary(
+        newt.newtId,
+        dedupedTargets
+    ); // for backward compatibility with old newt versions that don't support the new target format

    return {
        message: {
@@ -1,30 +1,20 @@
-import { db, ExitNode, newts, remoteExitNodes, Transaction } from "@server/db";
+import { db, newts, remoteExitNodes } from "@server/db";
 import { MessageHandler } from "@server/routers/ws";
 import { exitNodes, Newt, sites } from "@server/db";
 import { eq } from "drizzle-orm";
 import { addPeer, deletePeer } from "../gerbil/peers";
 import logger from "@server/logger";
 import config from "@server/lib/config";
-import { findNextAvailableCidr } from "@server/lib/ip";
 import {
+    ExitNodePingResult,
    selectBestExitNode,
    verifyExitNodeOrgAccess
 } from "#dynamic/lib/exitNodes";
+import { getUniqueSubnetForExitNode } from "@server/lib/exitNodes";
 import { fetchContainers } from "./dockerSocket";
-import { lockManager } from "#dynamic/lib/lock";
 import { buildTargetConfigurationForNewtClient } from "./buildConfiguration";
 import { canCompress } from "@server/lib/clientVersionChecks";

-export type ExitNodePingResult = {
-    exitNodeId: number;
-    latencyMs: number;
-    weight: number;
-    error?: string;
-    exitNodeName: string;
-    endpoint: string;
-    wasPreviouslyConnected: boolean;
-};
-
 export const handleNewtRegisterMessage: MessageHandler = async (context) => {
    const { message, client, sendToClient } = context;
    const newt = client as Newt;
@@ -94,9 +84,12 @@ export const handleNewtRegisterMessage: MessageHandler = async (context) => {
        fetchContainers(newt.newtId);
    }

-    let siteSubnet = oldSite.subnet;
+    let siteSubnet = oldSite.exitNodeSubnet;
    let exitNodeIdToQuery = oldSite.exitNodeId;
-    if (exitNodeId && (oldSite.exitNodeId !== exitNodeId || !oldSite.subnet)) {
+    if (
+        exitNodeId &&
+        (oldSite.exitNodeId !== exitNodeId || !oldSite.exitNodeSubnet)
+    ) {
        // This effectively moves the exit node to the new one
        exitNodeIdToQuery = exitNodeId; // Use the provided exitNodeId if it differs from the site's exitNodeId

@@ -115,7 +108,7 @@ export const handleNewtRegisterMessage: MessageHandler = async (context) => {
            return;
        }

-        const newSubnet = await getUniqueSubnetForSite(exitNode);
+        const newSubnet = await getUniqueSubnetForExitNode(exitNode);

        if (!newSubnet) {
            logger.error(
@@ -131,7 +124,7 @@ export const handleNewtRegisterMessage: MessageHandler = async (context) => {
            .set({
                pubKey: publicKey,
                exitNodeId: exitNodeId,
-                subnet: newSubnet
+                exitNodeSubnet: newSubnet
            })
            .where(eq(sites.siteId, siteId))
            .returning();
@@ -250,40 +243,3 @@ export const handleNewtRegisterMessage: MessageHandler = async (context) => {
        excludeSender: false // Include sender in broadcast
    };
 };
-
-async function getUniqueSubnetForSite(
-    exitNode: ExitNode,
-    trx: Transaction | typeof db = db
-): Promise<string | null> {
-    const lockKey = `subnet-allocation:${exitNode.exitNodeId}`;
-
-    return await lockManager.withLock(
-        lockKey,
-        async () => {
-            const sitesQuery = await trx
-                .select({
-                    subnet: sites.subnet
-                })
-                .from(sites)
-                .where(eq(sites.exitNodeId, exitNode.exitNodeId));
-
-            const blockSize = config.getRawConfig().gerbil.site_block_size;
-            const subnets = sitesQuery
-                .map((site) => site.subnet)
-                .filter(
-                    (subnet) =>
-                        subnet &&
-                        /^(\d{1,3}\.){3}\d{1,3}\/\d{1,2}$/.test(subnet)
-                )
-                .filter((subnet) => subnet !== null);
-            subnets.push(exitNode.address.replace(/\/\d+$/, `/${blockSize}`));
-            const newSubnet = findNextAvailableCidr(
-                subnets,
-                blockSize,
-                exitNode.address
-            );
-            return newSubnet;
-        },
-        5000 // 5 second lock TTL - subnet allocation should be quick
-    );
-}
@@ -5,7 +5,7 @@ export * from "./handleNewtRegisterMessage";
 export * from "./handleReceiveBandwidthMessage";
 export * from "./handleNewtGetConfigMessage";
 export * from "./handleSocketMessages";
-export * from "./handleNewtPingRequestMessage";
+export * from "./handleNewtExitNodesRequestMessage";
 export * from "./handleApplyBlueprintMessage";
 export * from "./handleNewtPingMessage";
 export * from "./handleNewtDisconnectingMessage";
@@ -19,6 +19,7 @@ import logger from "@server/logger";
 import { and, eq, inArray } from "drizzle-orm";
 import { addPeer, deletePeer } from "../newt/peers";
 import config from "@server/lib/config";
+import { SiR } from "react-icons/si";

 export async function buildSiteConfigurationForOlmClient(
    client: Client,
@@ -38,6 +39,8 @@ export async function buildSiteConfigurationForOlmClient(
        aliases: Alias[];
    }[] = [];

+    let exitNodeAliases: string[] = [];
+
    // Get all sites data
    const sitesData = await db
        .select()
@@ -48,10 +51,6 @@ export async function buildSiteConfigurationForOlmClient(
        )
        .where(eq(clientSitesAssociationsCache.clientId, client.clientId));

-    if (sitesData.length === 0) {
-        return siteConfigurations;
-    }
-
    // Batch-fetch every site resource this client has access to across ALL sites
    // in a single query, then group by siteId in memory. This avoids issuing one
    // query per site (which would be N round-trips for N sites).
@@ -68,8 +67,8 @@ export async function buildSiteConfigurationForOlmClient(
                clientSiteResourcesAssociationsCache.siteResourceId
            )
        )
-        .innerJoin(networks, eq(siteResources.networkId, networks.networkId))
-        .innerJoin(siteNetworks, eq(networks.networkId, siteNetworks.networkId))
+        .leftJoin(networks, eq(siteResources.networkId, networks.networkId))
+        .leftJoin(siteNetworks, eq(networks.networkId, siteNetworks.networkId))
        .where(
            and(
                eq(
@@ -81,7 +80,15 @@ export async function buildSiteConfigurationForOlmClient(
        );

    const siteResourcesBySiteId = new Map<number, SiteResource[]>();
+    let siteResourcesForExitNode = [];
    for (const row of allClientSiteResources) {
+        if (row.siteResource.requiresExitNodeConnection) {
+            siteResourcesForExitNode.push(row.siteResource);
+        }
+        if (!row.siteId) {
+            // because we are doing a leftJoin above to get the inference resources without a network / sites
+            continue;
+        }
        const arr = siteResourcesBySiteId.get(row.siteId);
        if (arr) {
            arr.push(row.siteResource);
@@ -90,6 +97,17 @@ export async function buildSiteConfigurationForOlmClient(
        }
    }

+    exitNodeAliases = siteResourcesForExitNode
+        .map((sr) => sr.fullDomain || sr.alias) // take either in case we introduce different resource types that don't have a fullDomain
+        .filter((a) => a != null);
+
+    if (sitesData.length == 0) {
+        return {
+            siteConfigurations,
+            exitNodeAliases
+        };
+    }
+
    // Batch-fetch exit nodes for all sites in one query (only needed in relay mode).
    const exitNodesById = new Map<number, typeof exitNodes.$inferSelect>();
    if (!jitMode && relay) {
@@ -167,7 +185,7 @@ export async function buildSiteConfigurationForOlmClient(
            peerOps.push(deletePeer(site.siteId, client.pubKey!));
        }

-        if (!site.subnet) {
+        if (!site.exitNodeSubnet) {
            logger.debug(`Site ${site.siteId} has no subnet, skipping`);
            continue;
        }
@@ -226,5 +244,8 @@ export async function buildSiteConfigurationForOlmClient(
        });
    }

-    return siteConfigurations;
+    return {
+        siteConfigurations,
+        exitNodeAliases
+    };
 }
@@ -0,0 +1,93 @@
+import { db, clients } from "@server/db";
+import { MessageHandler } from "@server/routers/ws";
+import { exitNodes, Olm } from "@server/db";
+import logger from "@server/logger";
+import { eq } from "drizzle-orm";
+import { listExitNodes } from "#dynamic/lib/exitNodes";
+import { calculateExitNodeWeight } from "@server/lib/exitNodes";
+
+export const handleOlmExitNodesRequestMessage: MessageHandler = async (
+    context
+) => {
+    const { message, client: olmClient, sendToClient } = context;
+    const olm = olmClient as Olm;
+
+    logger.info("Handling exit nodes request olm message!");
+
+    if (!olm) {
+        logger.warn("olm not found");
+        return;
+    }
+
+    // Get the olm's orgId through the client relationship
+    if (!olm.clientId) {
+        logger.warn("olm clientId not found");
+        return;
+    }
+
+    const [client] = await db
+        .select({ orgId: clients.orgId })
+        .from(clients)
+        .where(eq(clients.clientId, olm.clientId))
+        .limit(1);
+
+    if (!client || !client.orgId) {
+        logger.warn("client not found");
+        return;
+    }
+
+    const { noCloud, chainId } = message.data;
+
+    const exitNodesList = await listExitNodes(
+        client.orgId,
+        true,
+        noCloud || false,
+        olm.clientId
+    ); // filter for only the online ones
+
+    let lastExitNodeId = null;
+    if (olm.clientId) {
+        const [lastExitNode] = await db
+            .select()
+            .from(clients)
+            .where(eq(clients.clientId, olm.clientId))
+            .limit(1);
+        lastExitNodeId = lastExitNode?.exitNodeId || null;
+    }
+
+    const exitNodesPayload = await Promise.all(
+        exitNodesList.map(async (node) => {
+            const weight = await calculateExitNodeWeight(
+                node.exitNodeId,
+                node.maxConnections
+            );
+
+            if (weight === null) {
+                return null;
+            }
+
+            return {
+                exitNodeId: node.exitNodeId,
+                exitNodeName: node.name,
+                endpoint: node.endpoint,
+                weight,
+                wasPreviouslyConnected: node.exitNodeId === lastExitNodeId
+            };
+        })
+    );
+
+    // filter out null values
+    const filteredExitNodes = exitNodesPayload.filter((node) => node !== null);
+
+    return {
+        message: {
+            type: "olm/ping/exitNodes",
+            data: {
+                exitNodes: filteredExitNodes,
+                chainId: chainId
+            }
+        },
+        broadcast: false, // Send to all clients
+        excludeSender: false // Include sender in broadcast
+    };
+};
@@ -1,4 +1,4 @@
-import { db, orgs, primaryDb } from "@server/db";
+import { db, ExitNode, exitNodes, orgs, primaryDb } from "@server/db";
 import { MessageHandler } from "@server/routers/ws";
 import {
    clients,
@@ -22,6 +22,13 @@ import { canCompress } from "@server/lib/clientVersionChecks";
 import config from "@server/lib/config";
 import cache from "#dynamic/lib/cache"; // not using regional here because we need this in the register message handler before we know where the client is
 import { waitForClientRebuildIdle } from "@server/lib/rebuildClientAssociations";
+import {
+    ExitNodePingResult,
+    selectBestExitNode,
+    verifyExitNodeOrgAccess
+} from "#dynamic/lib/exitNodes";
+import { getUniqueSubnetForExitNode } from "@server/lib/exitNodes";
+import { addPeer, deletePeer } from "../gerbil/peers";

 const HOLEPUNCH_STALE_CHAIN_THRESHOLD = 18;
 const HOLEPUNCH_STALE_CHAIN_TTL_SECONDS = 1800;
@@ -49,11 +56,20 @@ export const handleOlmRegisterMessage: MessageHandler = async (context) => {
        olmAgent,
        orgId,
        userToken,
+        pingResults,
        fingerprint,
        postures,
+        backwardsCompatible,
        chainId
    } = message.data;

+    if (backwardsCompatible) {
+        logger.debug(
+            "[handleOlmRegisterMessage] Backwards compatible mode detected - not sending connect message and waiting for ping response."
+        );
+        return;
+    }
+
    if (!olm.clientId) {
        logger.warn("[handleOlmRegisterMessage] Olm client ID not found");
        sendOlmError(OlmErrorCodes.CLIENT_ID_NOT_FOUND, olm.olmId);
@@ -284,7 +300,64 @@ export const handleOlmRegisterMessage: MessageHandler = async (context) => {
        return;
    }

-    if (client.pubKey !== publicKey || client.archived) {
+    let exitNodeId: number | undefined;
+    if (pingResults) {
+        const bestPingResult = selectBestExitNode(
+            pingResults as ExitNodePingResult[]
+        );
+        if (!bestPingResult) {
+            logger.warn("No suitable exit node found based on ping results");
+        }
+        exitNodeId = bestPingResult?.exitNodeId;
+    }
+
+    let clientSubnet = client.exitNodeSubnet;
+    if (
+        exitNodeId &&
+        (client.exitNodeId !== exitNodeId || !client.exitNodeSubnet)
+    ) {
+        const { exitNode, hasAccess } = await verifyExitNodeOrgAccess(
+            exitNodeId,
+            client.orgId
+        );
+
+        if (!exitNode) {
+            logger.warn("[handleOlmRegisterMessage] Exit node not found", {
+                orgId: client.orgId,
+                clientId: client.clientId
+            });
+            return;
+        }
+
+        if (!hasAccess) {
+            logger.warn(
+                "[handleOlmRegisterMessage] Not authorized to use this exit node",
+                { orgId: client.orgId, clientId: client.clientId }
+            );
+            return;
+        }
+
+        // TODO: IF WE DO NOT HAVE AN INFERENCE RESOURCE DO WE NEED TO BE HOLDING A SUBNET ON THE CLIENT?
+
+        const newSubnet = await getUniqueSubnetForExitNode(exitNode);
+
+        if (!newSubnet) {
+            logger.error(
+                `[handleOlmRegisterMessage] No available subnets found for exit node id ${exitNodeId} and client id ${client.clientId}`,
+                { orgId: client.orgId, clientId: client.clientId }
+            );
+            return;
+        }
+
+        clientSubnet = newSubnet;
+    }
+
+    if (
+        client.pubKey !== publicKey ||
+        client.archived ||
+        client.exitNodeId !== exitNodeId ||
+        client.exitNodeSubnet !== clientSubnet
+    ) {
        logger.info(
            "[handleOlmRegisterMessage] Public key mismatch. Updating public key and clearing session info...",
            { orgId: client.orgId, clientId: client.clientId }
@@ -294,7 +367,9 @@ export const handleOlmRegisterMessage: MessageHandler = async (context) => {
            .update(clients)
            .set({
                pubKey: publicKey,
-                archived: false
+                archived: false,
+                exitNodeId: exitNodeId, // this can be undefined if no exit node was selected, which is fine just means we cant talk to the node or connect to it
+                exitNodeSubnet: clientSubnet
            })
            .where(eq(clients.clientId, client.clientId));

@@ -319,6 +394,24 @@ export const handleOlmRegisterMessage: MessageHandler = async (context) => {
            );
    }

+    if (client.pubKey && client.pubKey !== publicKey && client.exitNodeId) {
+        // test the old client to see if its different then remove
+        logger.info("Public key mismatch. Deleting old peer...");
+        await deletePeer(client.exitNodeId, client.pubKey);
+    }
+
+    if (clientSubnet && exitNodeId) {
+        try {
+            // add the peer to the exit node so it can connect
+            await addPeer(exitNodeId, {
+                publicKey: publicKey,
+                allowedIps: [clientSubnet]
+            });
+        } catch (error) {
+            logger.error(`Failed to add peer to exit node: ${error}`);
+        }
+    }
+
    let staleHolePunchChainCount: number | undefined;
    const hasChainId =
        chainId !== undefined && chainId !== null && String(chainId) !== "";
@@ -376,15 +469,29 @@ export const handleOlmRegisterMessage: MessageHandler = async (context) => {
        return;
    }

+    let exitNode: ExitNode | null = null;
+    if (exitNodeId) {
+        [exitNode] = await db
+            .select()
+            .from(exitNodes)
+            .where(eq(exitNodes.exitNodeId, exitNodeId))
+            .limit(1);
+    }
+
    // NOTE: its important that the client here is the old client and the public key is the new key
    await waitForClientRebuildIdle(olm.clientId);

-    const siteConfigurations = await buildSiteConfigurationForOlmClient(
-        client,
-        publicKey,
-        relay,
-        jitMode
-    );
+    const { siteConfigurations, exitNodeAliases } =
+        await buildSiteConfigurationForOlmClient(
+            client,
+            publicKey,
+            relay,
+            jitMode
+        );
+
+    // logger.info(
+    //     `ExitNode Aliases: ${exitNodeAliases}`
+    // );

    // Return connect message with all site configurations
    return {
@@ -394,6 +501,17 @@ export const handleOlmRegisterMessage: MessageHandler = async (context) => {
                sites: siteConfigurations,
                tunnelIP: client.subnet,
                utilitySubnet: org.utilitySubnet,
+                exitNode:
+                    exitNode && client.exitNodeSubnet
+                        ? {
+                              aliases: exitNodeAliases,
+                              connect: exitNodeAliases.length > 0, // we do not need to connect to the exit node if we do not have inference resources and right now all site resources on the exit node have an alias
+                              endpoint: `${exitNode.endpoint}:${exitNode.listenPort}`,
+                              publicKey: exitNode.publicKey,
+                              serverIP: exitNode.address.split("/")[0],
+                              tunnelIP: `${client.exitNodeSubnet.split("/")[0]}/${exitNode.address.split("/")[1]}` // we need to use the exit node's subnet mask here because the client will be using the exit node's subnet mask for its routing table so we can address it
+                          }
+                        : undefined,
                chainId: chainId
            }
        },
@@ -15,3 +15,4 @@ export * from "./handleOlmServerInitAddPeerHandshake";
 export * from "./offlineChecker";
 export * from "./handleOlmUnLocalMessage";
 export * from "./handleOlmLocalMessage";
+export * from "./handleOlmExitNodesRequestMessage";
@@ -1,6 +1,7 @@
 import {
    Client,
    db,
+    ExitNode,
    exitNodes,
    Olm,
    sites,
@@ -48,12 +49,26 @@ export async function sendOlmSyncMessage(olm: Olm, client: Client) {
    }

    // NOTE: WE ARE HARDCODING THE RELAY PARAMETER TO FALSE HERE BUT IN THE REGISTER MESSAGE ITS DEFINED BY THE CLIENT
-    const siteConfigurations = await buildSiteConfigurationForOlmClient(
-        client,
-        client.pubKey,
-        false,
-        jitMode
-    );
+    const { siteConfigurations, exitNodeAliases } =
+        await buildSiteConfigurationForOlmClient(
+            client,
+            client.pubKey,
+            false,
+            jitMode
+        );
+
+    // The exit node the client itself is assigned to (for site resources hosted
+    // on it, e.g. inference), same as what's sent in the initial olm/wg/connect
+    // message. This is separate from exitNodesData below, which is only the set
+    // of exit nodes used for hole punching to reach site peers.
+    let clientExitNode: ExitNode | null = null;
+    if (client.exitNodeId) {
+        [clientExitNode] = await db
+            .select()
+            .from(exitNodes)
+            .where(eq(exitNodes.exitNodeId, client.exitNodeId))
+            .limit(1);
+    }

    // Get all exit nodes from sites where the client has peers
    const clientSites = await db
@@ -113,11 +128,23 @@ export async function sendOlmSyncMessage(olm: Olm, client: Client) {
            type: "olm/sync",
            data: {
                sites: siteConfigurations,
-                exitNodes: exitNodesData
+                exitNodes: exitNodesData, // this is for the holepunch information
+                // this is for the backhaul connection to the exit node
+                exitNode:
+                    clientExitNode && client.exitNodeSubnet
+                        ? {
+                              aliases: exitNodeAliases,
+                              connect: exitNodeAliases.length > 0, // we do not need to connect to the exit node if we do not have inference resources and right now all site resources on the exit node have an alias
+                              endpoint: `${clientExitNode.endpoint}:${clientExitNode.listenPort}`,
+                              publicKey: clientExitNode.publicKey,
+                              serverIP: clientExitNode.address.split("/")[0],
+                              tunnelIP: client.exitNodeSubnet.split("/")[0]
+                          }
+                        : undefined
            }
        },
        {
-            compress: canCompress(olm.version, "olm")
+            compress: canCompress(olm.version, "olm") // we dont increment the version here or we could get into a loop!
        }
    ).catch((error) => {
        logger.warn(`Error sending olm sync message:`, error);
@@ -0,0 +1,154 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { db, resources, resourceAiModels } from "@server/db";
+import { eq, and } from "drizzle-orm";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import {
+    assertPublicModelListApiEligible,
+    assertPublicResourceModelEntriesValid,
+    modelListTypeSchema
+} from "@server/lib/aiInferenceResource";
+
+const addAiModelToResourceBodySchema = z.strictObject({
+    modelId: z.number().int().positive(),
+    listType: modelListTypeSchema.optional().default("allow")
+});
+
+const addAiModelToResourceParamsSchema = z.strictObject({
+    resourceId: z.coerce.number().int().positive()
+});
+
+registry.registerPath({
+    method: "post",
+    path: "/resource/{resourceId}/ai-models/add",
+    description:
+        "Add a single model to an inference resource allow/block selection. Requires at least one attached AI provider in select mode. The model must belong to a select-mode provider and its listType must match the provider catalog entry. listType defaults to allow.",
+    tags: [OpenAPITags.PublicResource],
+    request: {
+        params: addAiModelToResourceParamsSchema,
+        body: {
+            content: {
+                "application/json": {
+                    schema: addAiModelToResourceBodySchema
+                }
+            }
+        }
+    },
+    responses: {
+        200: {
+            description: "Successful response",
+            content: {
+                "application/json": {
+                    schema: z.object({
+                        data: z.record(z.string(), z.any()).nullable(),
+                        success: z.boolean(),
+                        error: z.boolean(),
+                        message: z.string(),
+                        status: z.number()
+                    })
+                }
+            }
+        }
+    }
+});
+
+export async function addAiModelToResource(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedBody = addAiModelToResourceBodySchema.safeParse(req.body);
+        if (!parsedBody.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedBody.error).toString()
+                )
+            );
+        }
+
+        const { modelId, listType } = parsedBody.data;
+
+        const parsedParams = addAiModelToResourceParamsSchema.safeParse(
+            req.params
+        );
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { resourceId } = parsedParams.data;
+
+        const [resource] = await db
+            .select()
+            .from(resources)
+            .where(eq(resources.resourceId, resourceId))
+            .limit(1);
+
+        if (!resource) {
+            return next(
+                createHttpError(HttpCode.NOT_FOUND, "Resource not found")
+            );
+        }
+
+        const eligibleError = await assertPublicModelListApiEligible(resource);
+        if (eligibleError) {
+            return next(createHttpError(HttpCode.BAD_REQUEST, eligibleError));
+        }
+
+        const modelError = await assertPublicResourceModelEntriesValid({
+            orgId: resource.orgId,
+            resourceId,
+            models: [{ modelId, listType }]
+        });
+        if (modelError) {
+            return next(createHttpError(HttpCode.BAD_REQUEST, modelError));
+        }
+
+        const existingEntry = await db
+            .select()
+            .from(resourceAiModels)
+            .where(
+                and(
+                    eq(resourceAiModels.resourceId, resourceId),
+                    eq(resourceAiModels.modelId, modelId)
+                )
+            );
+
+        if (existingEntry.length > 0) {
+            return next(
+                createHttpError(
+                    HttpCode.CONFLICT,
+                    "Model already assigned to resource"
+                )
+            );
+        }
+
+        await db
+            .insert(resourceAiModels)
+            .values({ resourceId, modelId, listType });
+
+        return response(res, {
+            data: {},
+            success: true,
+            error: false,
+            message: "Model added to resource successfully",
+            status: HttpCode.CREATED
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}
@@ -0,0 +1,158 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { db, resources } from "@server/db";
+import { eq } from "drizzle-orm";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import {
+    isInferenceFieldsError,
+    listPublicResourceAiProviders,
+    resolveProviderAttachments,
+    setPublicResourceAiProviders
+} from "@server/lib/aiInferenceResource";
+
+const addAiProviderToResourceBodySchema = z.strictObject({
+    providerId: z.number().int().positive()
+});
+
+const addAiProviderToResourceParamsSchema = z.strictObject({
+    resourceId: z.coerce.number().int().positive()
+});
+
+registry.registerPath({
+    method: "post",
+    path: "/resource/{resourceId}/ai-providers/add",
+    description:
+        "Add or replace a single AI provider attachment on an inference resource. The provider is attached in inherit mode, using its own allow/block lists.",
+    tags: [OpenAPITags.PublicResource],
+    request: {
+        params: addAiProviderToResourceParamsSchema,
+        body: {
+            content: {
+                "application/json": {
+                    schema: addAiProviderToResourceBodySchema
+                }
+            }
+        }
+    },
+    responses: {
+        200: {
+            description: "Successful response",
+            content: {
+                "application/json": {
+                    schema: z.object({
+                        data: z.record(z.string(), z.any()).nullable(),
+                        success: z.boolean(),
+                        error: z.boolean(),
+                        message: z.string(),
+                        status: z.number()
+                    })
+                }
+            }
+        }
+    }
+});
+
+export async function addAiProviderToResource(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedBody = addAiProviderToResourceBodySchema.safeParse(
+            req.body
+        );
+        if (!parsedBody.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedBody.error).toString()
+                )
+            );
+        }
+
+        const { providerId } = parsedBody.data;
+
+        const parsedParams = addAiProviderToResourceParamsSchema.safeParse(
+            req.params
+        );
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { resourceId } = parsedParams.data;
+
+        const [resource] = await db
+            .select()
+            .from(resources)
+            .where(eq(resources.resourceId, resourceId))
+            .limit(1);
+
+        if (!resource) {
+            return next(
+                createHttpError(HttpCode.NOT_FOUND, "Resource not found")
+            );
+        }
+
+        if (resource.mode !== "inference") {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    "AI providers can only be attached to inference-mode resources"
+                )
+            );
+        }
+
+        const existing = await listPublicResourceAiProviders(resourceId);
+        const nextAttachments = [
+            ...existing
+                .filter((a) => a.providerId !== providerId)
+                .map((a) => ({
+                    providerId: a.providerId,
+                    accessMode: a.accessMode,
+                    enabled: a.enabled
+                })),
+            {
+                providerId,
+                accessMode: "inherit" as const,
+                enabled: true as const
+            }
+        ];
+
+        const attachments = await resolveProviderAttachments({
+            orgId: resource.orgId,
+            attachments: nextAttachments,
+            requireAtLeastOne: true,
+            resourceId
+        });
+        if (isInferenceFieldsError(attachments)) {
+            return next(
+                createHttpError(HttpCode.BAD_REQUEST, attachments.error)
+            );
+        }
+
+        await setPublicResourceAiProviders(resourceId, attachments);
+
+        return response(res, {
+            data: {},
+            success: true,
+            error: false,
+            message: "AI provider added to resource successfully",
+            status: HttpCode.CREATED
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}
@@ -38,17 +38,24 @@ import {
 } from "@server/db/names";
 import { usageService } from "@server/lib/billing/usageService";
 import { LimitId } from "@server/lib/billing";
+import {
+    isInferenceFieldsError,
+    resolveProviderAttachments,
+    resourceAiProviderAttachmentSchema,
+    setPublicResourceAiProviders,
+    type ResourceAiProviderAttachment
+} from "@server/lib/aiInferenceResource";

 const createResourceParamsSchema = z.strictObject({
    orgId: z.string()
 });

 function resolveModeFromLegacyFields(data: {
-    mode?: "http" | "ssh" | "rdp" | "vnc" | "tcp" | "udp";
+    mode?: "http" | "ssh" | "rdp" | "vnc" | "tcp" | "udp" | "inference";
    http?: boolean;
    protocol?: "tcp" | "udp";
 }): {
-    mode?: "http" | "ssh" | "rdp" | "vnc" | "tcp" | "udp";
+    mode?: "http" | "ssh" | "rdp" | "vnc" | "tcp" | "udp" | "inference";
    error?: string;
 } {
    if (data.mode) {
@@ -90,11 +97,20 @@ const createHttpResourceSchema = z
        domainId: z.string(),
        stickySession: z.boolean().optional(),
        postAuthPath: z.string().nullable().optional(),
-        mode: z.enum(["http", "ssh", "rdp", "vnc", "tcp", "udp"]).optional(),
+        mode: z
+            .enum(["http", "ssh", "rdp", "vnc", "tcp", "udp", "inference"])
+            .optional(),
        // SSH Settings
        pamMode: z.enum(["passthrough", "push"]).optional(),
        authDaemonPort: z.int().positive().optional(),
-        authDaemonMode: z.enum(["site", "remote", "native"]).optional()
+        authDaemonMode: z.enum(["site", "remote", "native"]).optional(),
+        // Inference settings
+        aiProviders: z
+            .array(resourceAiProviderAttachmentSchema)
+            .optional()
+            .describe(
+                "For inference-mode resources: AI providers to attach. Providers are attached in inherit mode, using each provider's own allow/block lists. Effective allow model keys must be unique across attached providers."
+            )
    })
    .refine(
        (data) => {
@@ -365,11 +381,40 @@ async function createHttpResource(
        mode,
        authDaemonPort,
        authDaemonMode,
-        pamMode
+        pamMode,
+        aiProviders: aiProviderInputs
    } = parsedBody.data;
    const subdomain = parsedBody.data.subdomain;
    const stickySession = parsedBody.data.stickySession;

+    const effectiveMode = mode ?? "http";
+
+    let providerAttachments: ResourceAiProviderAttachment[] = [];
+    if (effectiveMode === "inference") {
+        // A new resource has no model selections yet, so providers always start
+        // in inherit mode; select can be enabled afterwards.
+        const resolved = await resolveProviderAttachments({
+            orgId,
+            attachments: (aiProviderInputs ?? []).map((p) => ({
+                providerId: p.providerId,
+                accessMode: "inherit" as const,
+                enabled: true as const
+            })),
+            requireAtLeastOne: false
+        });
+        if (isInferenceFieldsError(resolved)) {
+            return next(createHttpError(HttpCode.BAD_REQUEST, resolved.error));
+        }
+        providerAttachments = resolved;
+    } else if (aiProviderInputs && aiProviderInputs.length > 0) {
+        return next(
+            createHttpError(
+                HttpCode.BAD_REQUEST,
+                "AI providers can only be attached to inference-mode resources"
+            )
+        );
+    }
+
    // Wildcard subdomains are a paid feature
    if (subdomain && subdomain.includes("*")) {
        const isLicensed = await isLicensedOrSubscribed(
@@ -410,7 +455,7 @@ async function createHttpResource(
    }

    if (
-        ["ssh", "rdp", "vnc"].includes(mode!) &&
+        ["ssh", "rdp", "vnc"].includes(effectiveMode) &&
        !isLicensedOrSubscribed(
            orgId!,
            tierMatrix[TierFeature.AdvancedPublicResources]
@@ -543,7 +588,7 @@ async function createHttpResource(
                orgId,
                name,
                subdomain: finalSubdomain,
-                mode: mode,
+                mode: effectiveMode,
                pamMode: pamMode,
                authDaemonMode: authDaemonMode,
                authDaemonPort: authDaemonPort,
@@ -556,6 +601,14 @@ async function createHttpResource(
            })
            .returning();

+        if (providerAttachments.length > 0) {
+            await setPublicResourceAiProviders(
+                newResource[0].resourceId,
+                providerAttachments,
+                trx
+            );
+        }
+
        await trx.insert(roleResources).values({
            roleId: adminRole[0].roleId,
            resourceId: newResource[0].resourceId
@@ -35,3 +35,11 @@ export * from "./removeEmailFromResourceWhitelist";
 export * from "./getStatusHistory";
 export * from "./getBatchedStatusHistory";
 export * from "./getResourcePolicies";
+export * from "./listResourceAiModels";
+export * from "./setResourceAiModels";
+export * from "./addAiModelToResource";
+export * from "./removeAiModelFromResource";
+export * from "./listResourceAiProviders";
+export * from "./setResourceAiProviders";
+export * from "./addAiProviderToResource";
+export * from "./removeAiProviderFromResource";
@@ -0,0 +1,109 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { db, resources, resourceAiModels, aiModels } from "@server/db";
+import { eq } from "drizzle-orm";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+
+const listResourceAiModelsParamsSchema = z.strictObject({
+    resourceId: z.coerce.number().int().positive()
+});
+
+async function query(resourceId: number) {
+    return await db
+        .select({
+            modelId: aiModels.modelId,
+            modelKey: aiModels.modelKey,
+            name: aiModels.name,
+            providerId: aiModels.providerId,
+            enabled: aiModels.enabled,
+            listType: resourceAiModels.listType
+        })
+        .from(resourceAiModels)
+        .innerJoin(aiModels, eq(resourceAiModels.modelId, aiModels.modelId))
+        .where(eq(resourceAiModels.resourceId, resourceId));
+}
+
+export type ListResourceAiModelsResponse = {
+    models: NonNullable<Awaited<ReturnType<typeof query>>>;
+};
+
+registry.registerPath({
+    method: "get",
+    path: "/resource/{resourceId}/ai-models",
+    description:
+        "List the models this resource has selected from its select-mode providers' allow/block lists. Providers in inherit mode are not represented here; they use their own lists.",
+    tags: [OpenAPITags.PublicResource],
+    request: {
+        params: listResourceAiModelsParamsSchema
+    },
+    responses: {
+        200: {
+            description: "Successful response",
+            content: {
+                "application/json": {
+                    schema: z.object({
+                        data: z.record(z.string(), z.any()).nullable(),
+                        success: z.boolean(),
+                        error: z.boolean(),
+                        message: z.string(),
+                        status: z.number()
+                    })
+                }
+            }
+        }
+    }
+});
+
+export async function listResourceAiModels(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedParams = listResourceAiModelsParamsSchema.safeParse(
+            req.params
+        );
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { resourceId } = parsedParams.data;
+
+        const [resource] = await db
+            .select()
+            .from(resources)
+            .where(eq(resources.resourceId, resourceId))
+            .limit(1);
+
+        if (!resource) {
+            return next(
+                createHttpError(HttpCode.NOT_FOUND, "Resource not found")
+            );
+        }
+
+        const models = await query(resourceId);
+
+        return response<ListResourceAiModelsResponse>(res, {
+            data: { models },
+            success: true,
+            error: false,
+            message: "Resource AI models retrieved successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}
@@ -0,0 +1,95 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { db, resources } from "@server/db";
+import { eq } from "drizzle-orm";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { listPublicResourceAiProviders } from "@server/lib/aiInferenceResource";
+
+const listResourceAiProvidersParamsSchema = z.strictObject({
+    resourceId: z.coerce.number().int().positive()
+});
+
+export type ListResourceAiProvidersResponse = {
+    providers: Awaited<ReturnType<typeof listPublicResourceAiProviders>>;
+};
+
+registry.registerPath({
+    method: "get",
+    path: "/resource/{resourceId}/ai-providers",
+    description:
+        "List AI providers attached to an inference resource, including each attachment's accessMode.",
+    tags: [OpenAPITags.PublicResource],
+    request: {
+        params: listResourceAiProvidersParamsSchema
+    },
+    responses: {
+        200: {
+            description: "Successful response",
+            content: {
+                "application/json": {
+                    schema: z.object({
+                        data: z.record(z.string(), z.any()).nullable(),
+                        success: z.boolean(),
+                        error: z.boolean(),
+                        message: z.string(),
+                        status: z.number()
+                    })
+                }
+            }
+        }
+    }
+});
+
+export async function listResourceAiProviders(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedParams = listResourceAiProvidersParamsSchema.safeParse(
+            req.params
+        );
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { resourceId } = parsedParams.data;
+
+        const [resource] = await db
+            .select()
+            .from(resources)
+            .where(eq(resources.resourceId, resourceId))
+            .limit(1);
+
+        if (!resource) {
+            return next(
+                createHttpError(HttpCode.NOT_FOUND, "Resource not found")
+            );
+        }
+
+        const providers = await listPublicResourceAiProviders(resourceId);
+
+        return response<ListResourceAiProvidersResponse>(res, {
+            data: { providers },
+            success: true,
+            error: false,
+            message: "Resource AI providers retrieved successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}
@@ -637,11 +637,12 @@ export async function listResources(
                    ${resourcePassword.passwordId}
                )
            `;
-            const browserGatewayModes = ["http", "ssh", "rdp", "vnc"];
+            const browserGatewayModes = ["http", "ssh", "rdp", "vnc"] as const;

            switch (authState) {
                case "none":
                    conditions.push(
+                        // TODO: Does inference belong here?
                        or(eq(resources.mode, "tcp"), eq(resources.mode, "udp"))
                    );
                    break;
@@ -0,0 +1,147 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { db, resources, resourceAiModels } from "@server/db";
+import { eq, and } from "drizzle-orm";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { assertPublicModelListApiEligible } from "@server/lib/aiInferenceResource";
+
+const removeAiModelFromResourceBodySchema = z.strictObject({
+    modelId: z.int().positive()
+});
+
+const removeAiModelFromResourceParamsSchema = z.strictObject({
+    resourceId: z.coerce.number().int().positive()
+});
+
+registry.registerPath({
+    method: "post",
+    path: "/resource/{resourceId}/ai-models/remove",
+    description:
+        "Remove a single model from an inference resource allow/block list. Requires at least one attached AI provider.",
+    tags: [OpenAPITags.PublicResource],
+    request: {
+        params: removeAiModelFromResourceParamsSchema,
+        body: {
+            content: {
+                "application/json": {
+                    schema: removeAiModelFromResourceBodySchema
+                }
+            }
+        }
+    },
+    responses: {
+        200: {
+            description: "Successful response",
+            content: {
+                "application/json": {
+                    schema: z.object({
+                        data: z.record(z.string(), z.any()).nullable(),
+                        success: z.boolean(),
+                        error: z.boolean(),
+                        message: z.string(),
+                        status: z.number()
+                    })
+                }
+            }
+        }
+    }
+});
+
+export async function removeAiModelFromResource(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedBody = removeAiModelFromResourceBodySchema.safeParse(
+            req.body
+        );
+        if (!parsedBody.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedBody.error).toString()
+                )
+            );
+        }
+
+        const { modelId } = parsedBody.data;
+
+        const parsedParams = removeAiModelFromResourceParamsSchema.safeParse(
+            req.params
+        );
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { resourceId } = parsedParams.data;
+
+        const [resource] = await db
+            .select()
+            .from(resources)
+            .where(eq(resources.resourceId, resourceId))
+            .limit(1);
+
+        if (!resource) {
+            return next(
+                createHttpError(HttpCode.NOT_FOUND, "Resource not found")
+            );
+        }
+
+        const eligibleError = await assertPublicModelListApiEligible(resource);
+        if (eligibleError) {
+            return next(createHttpError(HttpCode.BAD_REQUEST, eligibleError));
+        }
+
+        const existingEntry = await db
+            .select()
+            .from(resourceAiModels)
+            .where(
+                and(
+                    eq(resourceAiModels.resourceId, resourceId),
+                    eq(resourceAiModels.modelId, modelId)
+                )
+            );
+
+        if (existingEntry.length === 0) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    "Model not found in resource's restriction list"
+                )
+            );
+        }
+
+        await db
+            .delete(resourceAiModels)
+            .where(
+                and(
+                    eq(resourceAiModels.resourceId, resourceId),
+                    eq(resourceAiModels.modelId, modelId)
+                )
+            );
+
+        return response(res, {
+            data: {},
+            success: true,
+            error: false,
+            message: "Model removed from resource successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}
@@ -0,0 +1,161 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { db, resources } from "@server/db";
+import { eq } from "drizzle-orm";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import {
+    isInferenceFieldsError,
+    listPublicResourceAiProviders,
+    resolveProviderAttachments,
+    setPublicResourceAiProviders
+} from "@server/lib/aiInferenceResource";
+
+const removeAiProviderFromResourceBodySchema = z.strictObject({
+    providerId: z.number().int().positive()
+});
+
+const removeAiProviderFromResourceParamsSchema = z.strictObject({
+    resourceId: z.coerce.number().int().positive()
+});
+
+registry.registerPath({
+    method: "post",
+    path: "/resource/{resourceId}/ai-providers/remove",
+    description:
+        "Remove an AI provider attachment from an inference resource. At least one provider must remain.",
+    tags: [OpenAPITags.PublicResource],
+    request: {
+        params: removeAiProviderFromResourceParamsSchema,
+        body: {
+            content: {
+                "application/json": {
+                    schema: removeAiProviderFromResourceBodySchema
+                }
+            }
+        }
+    },
+    responses: {
+        200: {
+            description: "Successful response",
+            content: {
+                "application/json": {
+                    schema: z.object({
+                        data: z.record(z.string(), z.any()).nullable(),
+                        success: z.boolean(),
+                        error: z.boolean(),
+                        message: z.string(),
+                        status: z.number()
+                    })
+                }
+            }
+        }
+    }
+});
+
+export async function removeAiProviderFromResource(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedBody = removeAiProviderFromResourceBodySchema.safeParse(
+            req.body
+        );
+        if (!parsedBody.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedBody.error).toString()
+                )
+            );
+        }
+
+        const { providerId } = parsedBody.data;
+
+        const parsedParams = removeAiProviderFromResourceParamsSchema.safeParse(
+            req.params
+        );
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { resourceId } = parsedParams.data;
+
+        const [resource] = await db
+            .select()
+            .from(resources)
+            .where(eq(resources.resourceId, resourceId))
+            .limit(1);
+
+        if (!resource) {
+            return next(
+                createHttpError(HttpCode.NOT_FOUND, "Resource not found")
+            );
+        }
+
+        if (resource.mode !== "inference") {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    "AI providers can only be attached to inference-mode resources"
+                )
+            );
+        }
+
+        const existing = await listPublicResourceAiProviders(resourceId);
+        const found = existing.find((a) => a.providerId === providerId);
+        if (!found) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    "AI provider is not attached to this resource"
+                )
+            );
+        }
+
+        const remaining = existing
+            .filter((a) => a.providerId !== providerId)
+            .map((a) => ({
+                providerId: a.providerId,
+                accessMode: a.accessMode,
+                enabled: a.enabled
+            }));
+
+        const attachments = await resolveProviderAttachments({
+            orgId: resource.orgId,
+            attachments: remaining,
+            requireAtLeastOne: false,
+            resourceId
+        });
+        if (isInferenceFieldsError(attachments)) {
+            return next(
+                createHttpError(HttpCode.BAD_REQUEST, attachments.error)
+            );
+        }
+
+        await setPublicResourceAiProviders(resourceId, attachments);
+
+        return response(res, {
+            data: {},
+            success: true,
+            error: false,
+            message: "AI provider removed from resource successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}
@@ -0,0 +1,153 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { db, resources, resourceAiModels } from "@server/db";
+import { eq } from "drizzle-orm";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import {
+    assertPublicModelListApiEligible,
+    assertPublicResourceModelEntriesValid,
+    resourceAiModelEntrySchema
+} from "@server/lib/aiInferenceResource";
+
+const setResourceAiModelsBodySchema = z.strictObject({
+    models: z.array(resourceAiModelEntrySchema)
+});
+
+const setResourceAiModelsParamsSchema = z.strictObject({
+    resourceId: z.coerce.number().int().positive()
+});
+
+registry.registerPath({
+    method: "post",
+    path: "/resource/{resourceId}/ai-models",
+    description:
+        "Replace the allow/block model selection for an inference resource. Requires at least one attached AI provider in select mode. Models must belong to a select-mode provider and their listType must match the provider catalog entry. An empty array clears the selection, which denies all models for select-mode providers.",
+    tags: [OpenAPITags.PublicResource],
+    request: {
+        params: setResourceAiModelsParamsSchema,
+        body: {
+            content: {
+                "application/json": {
+                    schema: setResourceAiModelsBodySchema
+                }
+            }
+        }
+    },
+    responses: {
+        200: {
+            description: "Successful response",
+            content: {
+                "application/json": {
+                    schema: z.object({
+                        data: z.record(z.string(), z.any()).nullable(),
+                        success: z.boolean(),
+                        error: z.boolean(),
+                        message: z.string(),
+                        status: z.number()
+                    })
+                }
+            }
+        }
+    }
+});
+
+export async function setResourceAiModels(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedBody = setResourceAiModelsBodySchema.safeParse(req.body);
+        if (!parsedBody.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedBody.error).toString()
+                )
+            );
+        }
+
+        const { models } = parsedBody.data;
+
+        const parsedParams = setResourceAiModelsParamsSchema.safeParse(
+            req.params
+        );
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { resourceId } = parsedParams.data;
+
+        const [resource] = await db
+            .select()
+            .from(resources)
+            .where(eq(resources.resourceId, resourceId))
+            .limit(1);
+
+        if (!resource) {
+            return next(
+                createHttpError(HttpCode.NOT_FOUND, "Resource not found")
+            );
+        }
+
+        const eligibleError = await assertPublicModelListApiEligible(resource);
+        if (eligibleError) {
+            return next(createHttpError(HttpCode.BAD_REQUEST, eligibleError));
+        }
+
+        const byModelId = new Map(
+            models.map((m) => [m.modelId, m.listType] as const)
+        );
+        const uniqueModels = [...byModelId.entries()].map(
+            ([modelId, listType]) => ({ modelId, listType })
+        );
+
+        const modelError = await assertPublicResourceModelEntriesValid({
+            orgId: resource.orgId,
+            resourceId,
+            models: uniqueModels
+        });
+        if (modelError) {
+            return next(createHttpError(HttpCode.BAD_REQUEST, modelError));
+        }
+
+        await db.transaction(async (trx) => {
+            await trx
+                .delete(resourceAiModels)
+                .where(eq(resourceAiModels.resourceId, resourceId));
+
+            if (uniqueModels.length > 0) {
+                await trx.insert(resourceAiModels).values(
+                    uniqueModels.map((m) => ({
+                        resourceId,
+                        modelId: m.modelId,
+                        listType: m.listType
+                    }))
+                );
+            }
+        });
+
+        return response(res, {
+            data: {},
+            success: true,
+            error: false,
+            message: "AI models set for resource successfully",
+            status: HttpCode.CREATED
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}
@@ -0,0 +1,140 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { db, resources } from "@server/db";
+import { eq } from "drizzle-orm";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import {
+    isInferenceFieldsError,
+    resolveProviderAttachments,
+    resourceAiProviderAttachmentSchema,
+    setPublicResourceAiProviders
+} from "@server/lib/aiInferenceResource";
+
+const setResourceAiProvidersBodySchema = z.strictObject({
+    providers: z.array(resourceAiProviderAttachmentSchema)
+});
+
+const setResourceAiProvidersParamsSchema = z.strictObject({
+    resourceId: z.coerce.number().int().positive()
+});
+
+registry.registerPath({
+    method: "post",
+    path: "/resource/{resourceId}/ai-providers",
+    description:
+        "Replace the AI providers attached to an inference resource. Each provider uses accessMode inherit (default, uses the provider's own allow/block lists) or select (uses the resource's selected subset of that provider's catalog). An empty list clears all providers. Effective allow model keys must be unique across attached providers.",
+    tags: [OpenAPITags.PublicResource],
+    request: {
+        params: setResourceAiProvidersParamsSchema,
+        body: {
+            content: {
+                "application/json": {
+                    schema: setResourceAiProvidersBodySchema
+                }
+            }
+        }
+    },
+    responses: {
+        200: {
+            description: "Successful response",
+            content: {
+                "application/json": {
+                    schema: z.object({
+                        data: z.record(z.string(), z.any()).nullable(),
+                        success: z.boolean(),
+                        error: z.boolean(),
+                        message: z.string(),
+                        status: z.number()
+                    })
+                }
+            }
+        }
+    }
+});
+
+export async function setResourceAiProviders(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedBody = setResourceAiProvidersBodySchema.safeParse(req.body);
+        if (!parsedBody.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedBody.error).toString()
+                )
+            );
+        }
+
+        const { providers } = parsedBody.data;
+
+        const parsedParams = setResourceAiProvidersParamsSchema.safeParse(
+            req.params
+        );
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { resourceId } = parsedParams.data;
+
+        const [resource] = await db
+            .select()
+            .from(resources)
+            .where(eq(resources.resourceId, resourceId))
+            .limit(1);
+
+        if (!resource) {
+            return next(
+                createHttpError(HttpCode.NOT_FOUND, "Resource not found")
+            );
+        }
+
+        if (resource.mode !== "inference") {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    "AI providers can only be attached to inference-mode resources"
+                )
+            );
+        }
+
+        const attachments = await resolveProviderAttachments({
+            orgId: resource.orgId,
+            attachments: providers,
+            requireAtLeastOne: false,
+            resourceId
+        });
+        if (isInferenceFieldsError(attachments)) {
+            return next(
+                createHttpError(HttpCode.BAD_REQUEST, attachments.error)
+            );
+        }
+
+        await setPublicResourceAiProviders(resourceId, attachments);
+
+        return response(res, {
+            data: {},
+            success: true,
+            error: false,
+            message: "AI providers set for resource successfully",
+            status: HttpCode.CREATED
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}
@@ -345,8 +345,10 @@ export async function updateResource(
            );
        }

-        if (["http", "ssh", "rdp", "vnc"].includes(resource.mode)) {
-            // HANDLE UPDATING HTTP RESOURCES
+        if (
+            ["http", "ssh", "rdp", "vnc", "inference"].includes(resource.mode)
+        ) {
+            // HANDLE UPDATING HTTP / BROWSER / INFERENCE RESOURCES
            return await updateHttpResource(
                {
                    req,
@@ -34,7 +34,9 @@ const createRoleSchema = z.strictObject({
 export const defaultRoleAllowedActions: ActionsEnum[] = [
    ActionsEnum.getOrg,
    ActionsEnum.getResource,
-    ActionsEnum.listResources
+    ActionsEnum.listResources,
+    ActionsEnum.getSiteResource,
+    ActionsEnum.listSiteResources
 ];

 export type CreateRoleBody = z.infer<typeof createRoleSchema>;
@@ -311,13 +311,13 @@ export async function createSite(
                // lets also make sure there is no overlap with other sites on the exit node
                const sitesQuery = await db
                    .select({
-                        subnet: sites.subnet
+                        subnet: sites.exitNodeSubnet
                    })
                    .from(sites)
                    .where(
                        and(
                            eq(sites.exitNodeId, exitNodeId),
-                            eq(sites.subnet, subnet)
+                            eq(sites.exitNodeSubnet, subnet)
                        )
                    );

@@ -427,7 +427,7 @@ export async function createSite(
                            exitNodeId,
                            name,
                            niceId: updatedNiceId!,
-                            subnet,
+                            exitNodeSubnet: subnet,
                            type,
                            pubKey: pubKey || null,
                            status: "approved"
@@ -444,7 +444,7 @@ export async function createSite(
                            type,
                            dockerSocketEnabled: false,
                            online: true,
-                            subnet: "0.0.0.0/32",
+                            exitNodeSubnet: "0.0.0.0/32",
                            status: "approved"
                        })
                        .returning();
@@ -125,7 +125,7 @@ function querySitesBase() {
            niceId: sites.niceId,
            name: sites.name,
            pubKey: sites.pubKey,
-            subnet: sites.subnet,
+            subnet: sites.exitNodeSubnet,
            megabytesIn: sites.megabytesIn,
            megabytesOut: sites.megabytesOut,
            orgName: orgs.name,
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
miloschwartz	9eafa067b9	improved provider picker ui on resource	2026-08-07 16:13:51 -04:00
miloschwartz	5e5e172d39	add streaming function to capability	2026-08-07 14:36:15 -04:00
miloschwartz	bc7a883f6c	add some parallelization to ai gateway pipeline	2026-08-07 14:36:11 -04:00
Owen	12056aebc6	Add basic cost calculations for testing	2026-08-07 13:54:32 -04:00
miloschwartz	297cb9c8f2	set provider default capabilities	2026-08-07 12:25:29 -04:00
miloschwartz	d699455b38	support changing capabilities on provider integrations	2026-08-07 11:23:07 -04:00
miloschwartz	9e7b4afaec	support allow list and blocklist	2026-08-07 10:59:50 -04:00
Owen	07f628b928	add the pangolin header information	2026-08-06 17:53:51 -04:00
Owen	184e1425a4	Private connection working with traefik	2026-08-06 17:22:01 -04:00
Owen	22f2990f56	support streaming and closes properly on site targets	2026-08-06 17:15:11 -04:00
Owen	6c28c5f383	Basic target routing	2026-08-06 15:48:13 -04:00
Owen	e9f9cf54f4	Include the sheme in the routing header	2026-08-06 15:18:11 -04:00
Owen	751d1b083d	Add back port restrictions	2026-08-06 14:55:07 -04:00
Owen	a4d77a4fd3	Fix info box ui	2026-08-06 14:45:25 -04:00
Owen	f72252552f	Fix saving	2026-08-06 14:28:14 -04:00
miloschwartz	c5d68675c9	support patterns in model key	2026-08-06 14:27:08 -04:00
miloschwartz	36b8ef5fba	Merge branch 'dev' into aig	2026-08-06 13:59:07 -04:00
Owen	6cca5e0472	Use fulldoman instead of the alias for the site resources	2026-08-06 13:45:51 -04:00
miloschwartz	4048fa274a	fix non admins cant see private resources details in launcher	2026-08-06 12:32:19 -04:00
Owen	b4d2974e19	Properly configure ssl and domain when creating and editing	2026-08-06 12:20:56 -04:00
Owen	fb896d6c0f	Add domain to the resource selection	2026-08-06 12:03:40 -04:00
miloschwartz	82b86263dc	allow chars in 2fa input form closes #3532	2026-08-06 11:39:06 -04:00
Owen	6564bfe8ae	disabled advanced mode on the providers	2026-08-06 11:00:34 -04:00
Owen	75ce7e91d7	Restrict the ai site resource router to exit node subnet	2026-08-06 11:00:34 -04:00
Owen	656eea5bb1	add a target routing provider for custom target resources	2026-08-06 11:00:34 -04:00
miloschwartz	aba27a7bbf	support skip tls per request	2026-08-05 17:53:28 -04:00
miloschwartz	3d7e322bf9	add headers to provider	2026-08-05 17:42:58 -04:00
miloschwartz	1f3fff4a9d	Merge branch 'aig' of https://github.com/fosrl/pangolin into aig	2026-08-05 16:55:58 -04:00
miloschwartz	39e06f2b6d	add api capabilities	2026-08-05 16:55:48 -04:00
Owen	b775c5b674	include clients in the peer config	2026-08-05 16:24:09 -04:00
Owen	796bf37141	ai gateway override only required sometimes	2026-08-05 16:24:01 -04:00
Owen	c5b0e1e876	add QUERY_LOGGING env to sqlite driver for verbose logging	2026-08-05 16:17:28 -04:00
miloschwartz	2e8bd7a8c7	improve form error codes	2026-08-05 15:56:12 -04:00
miloschwartz	790daba796	allow no provider on a resource	2026-08-05 15:43:42 -04:00
miloschwartz	bcf6b86b84	add no auth and passthrough auth	2026-08-05 15:39:08 -04:00
miloschwartz	2e9bd50172	add provider specific auth modes	2026-08-05 15:17:50 -04:00
miloschwartz	c673dce484	pin docker base image	2026-08-05 13:34:45 -04:00
Owen	e1dd8965dc	Remove extra checks	2026-08-05 11:31:12 -04:00
Owen	346a0bdc98	Resolve syntax issue	2026-08-05 10:31:14 -04:00
Owen	8c1169738e	Merge branch 'aig' of github.com:fosrl/pangolin into aig	2026-08-05 10:24:42 -04:00
miloschwartz	bc80f91a45	gateway endpoint mvp	2026-08-05 10:24:37 -04:00
Owen	3d062389e9	update oss with gateway traefik routers	2026-08-05 10:14:40 -04:00
Owen	7afddb5eb5	Clean logging	2026-08-05 10:00:36 -04:00
Owen	425a99e5ee	Show the auth settings for testing	2026-08-05 09:59:41 -04:00
Owen	a5e9339af9	Merge branch 'dev' into aig	2026-08-05 09:32:03 -04:00
Owen	835a30cffe	Show the cert status of the namespace domains properly	2026-08-04 17:44:34 -04:00
miloschwartz	83e20c2dfa	adjust placeholder ui for rules and targets	2026-08-04 17:36:00 -04:00
miloschwartz	f85d41945a	Merge branch 'aig' of https://github.com/fosrl/pangolin into aig	2026-08-04 17:27:10 -04:00
miloschwartz	149eb17b27	create basic public inference resource	2026-08-04 17:26:57 -04:00
Owen	18b90da6ab	Merge branch 'main' into dev	2026-08-04 17:22:32 -04:00
Owen	72d469b19c	Revert changes to traefik config to pull site resources	2026-08-04 17:17:51 -04:00
Owen	6526d7f225	Remove completions endpoints	2026-08-04 17:09:44 -04:00
Owen	36c0edc62e	Use p host to pass the host header	2026-08-04 17:08:59 -04:00
Owen	aad26b9ae4	move the ai gateway to its own server	2026-08-04 17:08:59 -04:00
Owen	80dcdfe251	change override to work for bother badger and ai gateway	2026-08-04 17:08:59 -04:00
miloschwartz	1696fc37a8	dont set port restrictions for inference	2026-08-04 17:04:29 -04:00
miloschwartz	7759d87835	add basic ui for private inference resource	2026-08-04 16:54:23 -04:00
miloschwartz	c085de1e9e	Merge branch 'aig' of https://github.com/fosrl/pangolin into aig Co-authored-by: Cursor <cursoragent@cursor.com>	2026-08-04 15:56:27 -04:00
miloschwartz	e38359c74f	add crud for adding providers and models to resources	2026-08-04 15:54:24 -04:00
Owen	2cfd7e867b	delete the client peers from the exit nodes properly	2026-08-04 15:46:22 -04:00
Owen	ec5a2b0cbe	add peers to exit node	2026-08-04 15:46:22 -04:00
Owen	2bf426bc22	dont restrict to enterprise and dont link with providers yet	2026-08-04 15:46:21 -04:00
miloschwartz	ed8545f8a2	update schema to support models and provider access on inferences resources and add budgets	2026-08-04 14:50:07 -04:00
Owen	9811492a0b	send the aliases when connecting for the first time	2026-08-04 14:23:25 -04:00
Owen	973925b35d	allow creating basic inference resource	2026-08-04 14:22:36 -04:00
Owen	0b30cfc341	Working on inference resource management	2026-08-04 11:48:36 -04:00
Owen	a7e44944fb	Use ip to find user	2026-08-04 10:43:49 -04:00
Owen	9b25ebd6cd	remove token fallback	2026-08-04 10:19:22 -04:00
miloschwartz	33b775e263	clean up providers ui	2026-08-04 10:15:49 -04:00
miloschwartz	b0edc6302e	providers table, create, and edit first pass	2026-08-04 10:15:49 -04:00
Owen	1a644b131d	Pull the session cookie properly	2026-08-04 10:15:49 -04:00
Owen	f7689b7a5a	Endpoints to update models on the resource	2026-08-04 10:15:49 -04:00
Owen	1073011a2a	first pass of traefik -> basic gateway	2026-08-04 10:15:49 -04:00
miloschwartz	6fa0009ebf	add targets and refactor endpoints	2026-08-04 10:15:49 -04:00
Owen	42c0abedb7	Add basic page to configure the alias	2026-08-04 10:15:48 -04:00
Owen	56e59a93f3	Handle the alias create and update special case exit nodes	2026-08-04 10:15:48 -04:00
Owen	9d581f3897	Fix types	2026-08-04 10:15:48 -04:00
miloschwartz	e5ac6ec7cd	add default providers and support overrides	2026-08-04 10:15:48 -04:00
miloschwartz	694fe98131	add crud for providers and models	2026-08-04 10:15:48 -04:00
miloschwartz	730078597e	add basic schema	2026-08-04 10:15:48 -04:00
Owen	32ac8db803	Handle aliases when registering	2026-08-04 10:15:48 -04:00
Owen	aad2ed2719	Get exit node every time	2026-08-04 10:15:48 -04:00
Owen	a790fed297	sync the exit node connection	2026-08-04 10:15:48 -04:00
Owen	deb2d5ce2a	Update site resources to handle new inference	2026-08-04 10:15:48 -04:00
Owen	093097c619	send if we need to connect to the gerbil or not	2026-08-04 10:15:48 -04:00
Owen	2cdeb7c104	Add inference resource type	2026-08-04 10:15:47 -04:00
Owen	4dbb04bfb8	Handle the ping exit node request backward mode for olm and install handler	2026-08-04 10:15:47 -04:00
Owen	fa8b921635	Fix postgres schema exit node subnet	2026-08-04 10:15:47 -04:00
Owen	33dd10c670	Rename subnet for clarity, pick subnet on client	2026-08-04 10:15:47 -04:00
Owen	ba24e1c4f5	Add exit node selection to the clients	2026-08-04 10:15:47 -04:00
Owen Schwartz	efe22c889c	Merge pull request #3522 from fosrl/dev Move rate linmit to file	2026-08-03 14:35:43 -04:00