Pull the session cookie properly

feat: show the correct idp type icon in the last used idp

messages/en-US.json

@@ -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",
@@ -2246,6 +2256,7 @@
     "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 +2412,7 @@
     "editInternalResourceDialogModeCidr": "CIDR",
     "editInternalResourceDialogModeHttp": "HTTP",
     "editInternalResourceDialogModeHttps": "HTTPS",
+    "editInternalResourceDialogModeInference": "Inference",
     "editInternalResourceDialogModeSsh": "SSH",
     "editInternalResourceDialogScheme": "Scheme",
     "editInternalResourceDialogEnableSsl": "Enable TLS",

server/auth/actions.ts

@@ -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(

server/db/pg/schema/schema.ts

@@ -99,6 +99,7 @@ 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
         megabytesIn: real("bytesIn").default(0),
         megabytesOut: real("bytesOut").default(0),
         lastBandwidthUpdate: varchar("lastBandwidthUpdate"),
@@ -194,7 +195,12 @@ 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"),
@@ -204,7 +210,11 @@ export const resources = pgTable(
         authDaemonPort: integer("authDaemonPort").default(22123),
         status: varchar("status")
             .$type<"pending" | "approved">()
-            .default("approved")
+            .default("approved"),
+        aiProviderId: integer("aiProviderId").references(
+            () => aiProviders.providerId,
+            { onDelete: "set null" }
+        )
     },
     (t) => [
         index("idx_resources_fulldomain")
@@ -215,6 +225,19 @@ export const resources = pgTable(
     ]
 );
 
+export const resourceAiModels = pgTable(
+    "resourceAiModels",
+    {
+        resourceId: integer("resourceId")
+            .notNull()
+            .references(() => resources.resourceId, { onDelete: "cascade" }),
+        modelId: integer("modelId")
+            .notNull()
+            .references(() => aiModels.modelId, { onDelete: "cascade" })
+    },
+    (t) => [primaryKey({ columns: [t.resourceId, t.modelId] })]
+);
+
 export const labels = pgTable("labels", {
     labelId: serial("labelId").primaryKey(),
     name: varchar("name").notNull(),
@@ -317,11 +340,18 @@ export const targets = pgTable(
     "targets",
     {
         targetId: serial("targetId").primaryKey(),
-        resourceId: integer("resourceId")
+        resourceId: integer("resourceId").references(
-            .references(() => resources.resourceId, {
+            () => resources.resourceId,
+            {
                 onDelete: "cascade"
-            })
+            }
-            .notNull(),
+        ),
+        providerId: integer("providerId").references(
+            () => aiProviders.providerId,
+            {
+                onDelete: "cascade"
+            }
+        ),
         siteId: integer("siteId")
             .references(() => sites.siteId, {
                 onDelete: "cascade"
@@ -345,6 +375,7 @@ 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`${t.enabled} = true`)
@@ -424,11 +455,14 @@ export const siteResources = pgTable(
                 onDelete: "restrict"
             }
         ),
+        requiresExitNodeConnection: boolean("requiresExitNodeConnection")
+            .notNull()
+            .default(false),
         niceId: varchar("niceId").notNull(),
         name: varchar("name").notNull(),
         ssl: boolean("ssl").notNull().default(false),
         mode: varchar("mode")
-            .$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
@@ -458,11 +492,30 @@ export const siteResources = pgTable(
         fullDomain: varchar("fullDomain"),
         status: varchar("status")
             .$type<"pending" | "approved">()
-            .default("approved")
+            .default("approved"),
+        aiProviderId: integer("aiProviderId").references(
+            () => aiProviders.providerId,
+            { onDelete: "set null" }
+        )
     },
     (t) => [index("idx_siteresources_orgid_niceid").on(t.orgId, t.niceId)]
 );
 
+export const siteResourceAiModels = pgTable(
+    "siteResourceAiModels",
+    {
+        siteResourceId: integer("siteResourceId")
+            .notNull()
+            .references(() => siteResources.siteResourceId, {
+                onDelete: "cascade"
+            }),
+        modelId: integer("modelId")
+            .notNull()
+            .references(() => aiModels.modelId, { onDelete: "cascade" })
+    },
+    (t) => [primaryKey({ columns: [t.siteResourceId, t.modelId] })]
+);
+
 export const networks = pgTable(
     "networks",
     {
@@ -1181,7 +1234,7 @@ 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(),
         megabytesIn: real("bytesIn"),
         megabytesOut: real("bytesOut"),
         lastBandwidthUpdate: varchar("lastBandwidthUpdate"),
@@ -1540,6 +1593,61 @@ 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">(),
+    routingMode: varchar("routingMode")
+        .$type<"url" | "target">()
+        .notNull()
+        .default("url"),
+    skipTlsVerification: boolean("skipTlsVerification")
+        .notNull()
+        .default(false),
+    budgetAmount: real("budgetAmount"),
+    budgetUnit: varchar("budgetUnit").$type<"usd" | "tokens">(),
+    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(),
+        budgetAmount: real("budgetAmount"),
+        budgetUnit: varchar("budgetUnit").$type<"usd" | "tokens">(),
+        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 type Org = InferSelectModel<typeof orgs>;
 export type User = InferSelectModel<typeof users>;
 export type Site = InferSelectModel<typeof sites>;
@@ -1624,3 +1732,7 @@ 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 ResourceAiModel = InferSelectModel<typeof resourceAiModels>;
+export type SiteResourceAiModel = InferSelectModel<typeof siteResourceAiModels>;

server/db/sqlite/schema/schema.ts

@@ -4,6 +4,7 @@ import {
     index,
     integer,
     primaryKey,
+    real,
     sqliteTable,
     text,
     unique
@@ -107,7 +108,7 @@ 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 +204,10 @@ 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"),
@@ -211,9 +215,26 @@ export const resources = sqliteTable("resources", {
         .$type<"site" | "remote" | "native">()
         .default("site"),
     authDaemonPort: integer("authDaemonPort").default(22123),
-    status: text("status").$type<"pending" | "approved">().default("approved")
+    status: text("status").$type<"pending" | "approved">().default("approved"),
+    aiProviderId: integer("aiProviderId").references(
+        () => aiProviders.providerId,
+        { onDelete: "set null" }
+    )
 });
 
+export const resourceAiModels = sqliteTable(
+    "resourceAiModels",
+    {
+        resourceId: integer("resourceId")
+            .notNull()
+            .references(() => resources.resourceId, { onDelete: "cascade" }),
+        modelId: integer("modelId")
+            .notNull()
+            .references(() => aiModels.modelId, { onDelete: "cascade" })
+    },
+    (t) => [primaryKey({ columns: [t.resourceId, t.modelId] })]
+);
+
 export const labels = sqliteTable("labels", {
     labelId: integer("labelId").primaryKey({ autoIncrement: true }),
     name: text("name").notNull(),
@@ -322,11 +343,12 @@ export const clientLabels = sqliteTable(
 
 export const targets = sqliteTable("targets", {
     targetId: integer("targetId").primaryKey({ autoIncrement: true }),
-    resourceId: integer("resourceId")
+    resourceId: integer("resourceId").references(() => resources.resourceId, {
-        .references(() => resources.resourceId, {
+        onDelete: "cascade"
-            onDelete: "cascade"
+    }),
-        })
+    providerId: integer("providerId").references(() => aiProviders.providerId, {
-        .notNull(),
+        onDelete: "cascade"
+    }),
     siteId: integer("siteId")
         .references(() => sites.siteId, {
             onDelete: "cascade"
@@ -422,10 +444,17 @@ 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").$type<"http" | "https">(), // only for when we are doing https or http mode
     proxyPort: integer("proxyPort"), // only for port mode
     destinationPort: integer("destinationPort"), // only for port mode
@@ -450,9 +479,28 @@ export const siteResources = sqliteTable("siteResources", {
     }),
     subdomain: text("subdomain"),
     fullDomain: text("fullDomain"),
-    status: text("status").$type<"pending" | "approved">().default("approved")
+    status: text("status").$type<"pending" | "approved">().default("approved"),
+    aiProviderId: integer("aiProviderId").references(
+        () => aiProviders.providerId,
+        { onDelete: "set null" }
+    )
 });
 
+export const siteResourceAiModels = sqliteTable(
+    "siteResourceAiModels",
+    {
+        siteResourceId: integer("siteResourceId")
+            .notNull()
+            .references(() => siteResources.siteResourceId, {
+                onDelete: "cascade"
+            }),
+        modelId: integer("modelId")
+            .notNull()
+            .references(() => aiModels.modelId, { onDelete: "cascade" })
+    },
+    (t) => [primaryKey({ columns: [t.siteResourceId, t.modelId] })]
+);
+
 export const networks = sqliteTable("networks", {
     networkId: integer("networkId").primaryKey({ autoIncrement: true }),
     niceId: text("niceId"),
@@ -599,6 +647,7 @@ 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 +1575,63 @@ 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">(),
+    routingMode: text("routingMode")
+        .$type<"url" | "target">()
+        .notNull()
+        .default("url"),
+    skipTlsVerification: integer("skipTlsVerification", { mode: "boolean" })
+        .notNull()
+        .default(false),
+    budgetAmount: real("budgetAmount"),
+    budgetUnit: text("budgetUnit").$type<"usd" | "tokens">(),
+    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(),
+        budgetAmount: real("budgetAmount"),
+        budgetUnit: text("budgetUnit").$type<"usd" | "tokens">(),
+        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 type Org = InferSelectModel<typeof orgs>;
 export type User = InferSelectModel<typeof users>;
 export type Site = InferSelectModel<typeof sites>;
@@ -1608,3 +1714,7 @@ 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 ResourceAiModel = InferSelectModel<typeof resourceAiModels>;
+export type SiteResourceAiModel = InferSelectModel<typeof siteResourceAiModels>;

server/index.ts

@@ -9,6 +9,8 @@ import { createIntegrationApiServer } from "./integrationApiServer";
 import {
     ApiKey,
     ApiKeyOrg,
+    AiModel,
+    AiProvider,
     RemoteExitNode,
     Session,
     SiteResource,
@@ -83,6 +85,8 @@ declare global {
             userOrgIds?: string[];
             remoteExitNode?: RemoteExitNode;
             siteResource?: SiteResource;
+            aiProvider?: AiProvider;
+            aiModel?: AiModel;
             orgPolicyAllowed?: boolean;
         }
     }

server/lib/aiProviderDefaults.ts

@@ -0,0 +1,106 @@
+export type AiProviderType =
+    | "openai"
+    | "anthropic"
+    | "googleGemini"
+    | "vertexAi"
+    | "bedrock"
+    | "microsoftFoundry"
+    | "openRouter"
+    | "vercelAiGateway"
+    | "custom";
+
+export type AiProviderAuthType = "bearer";
+export type AiBudgetUnit = "usd" | "tokens";
+export type AiProviderRoutingMode = "url" | "target";
+
+type AiProviderDefaults = {
+    upstreamUrl: string | null;
+    authType: AiProviderAuthType;
+};
+
+export const AI_PROVIDER_DEFAULTS: Record<
+    Exclude<AiProviderType, "custom">,
+    AiProviderDefaults
+> = {
+    openai: {
+        upstreamUrl: "https://api.openai.com/v1",
+        authType: "bearer"
+    },
+    anthropic: {
+        upstreamUrl: "https://api.anthropic.com",
+        authType: "bearer"
+    },
+    googleGemini: {
+        upstreamUrl: "https://generativelanguage.googleapis.com/v1beta/openai/",
+        authType: "bearer"
+    },
+    vertexAi: {
+        upstreamUrl: null,
+        authType: "bearer"
+    },
+    bedrock: {
+        upstreamUrl: "https://bedrock-runtime.us-east-1.amazonaws.com",
+        authType: "bearer"
+    },
+    microsoftFoundry: {
+        upstreamUrl: null,
+        authType: "bearer"
+    },
+    openRouter: {
+        upstreamUrl: "https://openrouter.ai/api/v1",
+        authType: "bearer"
+    },
+    vercelAiGateway: {
+        upstreamUrl: "https://ai-gateway.vercel.sh/v1",
+        authType: "bearer"
+    }
+};
+
+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 resolveAiProviderConfig(input: {
+    type: AiProviderType;
+    upstreamUrl: string | null;
+    authType: AiProviderAuthType | null;
+    routingMode?: AiProviderRoutingMode | null;
+}): {
+    upstreamUrl: string | null;
+    authType: AiProviderAuthType | null;
+    routingMode: AiProviderRoutingMode;
+} {
+    const routingMode = input.routingMode ?? "url";
+
+    if (routingMode === "target") {
+        return {
+            upstreamUrl: null,
+            authType: input.authType ?? "bearer",
+            routingMode
+        };
+    }
+
+    if (input.type === "custom") {
+        return {
+            upstreamUrl: input.upstreamUrl,
+            authType: input.authType,
+            routingMode
+        };
+    }
+
+    const defaults = AI_PROVIDER_DEFAULTS[input.type];
+    return {
+        upstreamUrl: input.upstreamUrl ?? defaults.upstreamUrl,
+        authType: input.authType ?? defaults.authType,
+        routingMode
+    };
+}

server/lib/alerts/events/healthCheckEvents.ts

@@ -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(
+    const allHealthy = monitoredTargets.every((t) => t.hcHealth === "healthy");
-        (t) => t.hcHealth === "healthy"
-    );
     const allUnhealthy = monitoredTargets.every(
         (t) => t.hcHealth === "unhealthy"
     );

server/lib/calculateUserClientsForOrgs.ts

@@ -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,

server/lib/deleteResource.ts

@@ -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[]>();

server/lib/deleteSiteAssociatedResources.ts

@@ -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(

server/lib/exitNodes/calculateExitNodeWeight.ts

@@ -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;
+}

server/lib/exitNodes/exitNodes.ts

@@ -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 {

server/lib/exitNodes/getUniqueSubnetForExitNode.ts

@@ -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
+    );
+}

server/lib/exitNodes/index.ts

@@ -2,3 +2,5 @@ export * from "./exitNodes";
 export * from "./exitNodeComms";
 export * from "./subnet";
 export * from "./getCurrentExitNodeId";
+export * from "./calculateExitNodeWeight";
+export * from "./getUniqueSubnetForExitNode";

server/lib/ip.ts

@@ -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(
+export function dedupeCertsForTargets(targetsV2: SubnetProxyTargetV2[]): {
-    targetsV2: SubnetProxyTargetV2[]
+    targets: SubnetProxyTargetV2[];
-): { targets: SubnetProxyTargetV2[]; certs: CertRef[] } {
+    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 {

server/lib/rebuildClientAssociations.ts

@@ -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 +1076,236 @@ 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
+        })
+        .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)
+    );
+
+    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 };
+    }[] = [];
+    const disconnectPayloads: {
+        clientId: string;
+        message: { type: string; data: any };
+        options: { compress: 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]
+                    }
+                },
+                options: { compress: canCompress(olm.version, "olm") }
+            });
+        } else {
+            disconnectPayloads.push({
+                clientId: olm.olmId,
+                message: {
+                    type: "olm/wg/exitnode/disconnect",
+                    data: {}
+                },
+                options: { compress: canCompress(olm.version, "olm") }
+            });
+        }
+    }
+
+    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") }
+        }));
+
+    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 +1536,7 @@ 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 +1963,38 @@ 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 +2276,10 @@ 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(

server/lib/traefik/TraefikConfigManager.ts

@@ -516,6 +516,9 @@ export class TraefikConfigManager {
             const maintenanceHost =
                 config.getRawConfig().server.internal_hostname;
             const pangolinUIUrl = `http://${maintenanceHost}:${maintenancePort}`;
+            const aiGatewayUrl = `http://${maintenanceHost}:${
+                config.getRawConfig().server.internal_port
+            }/api/v1/ai-gateway`;
 
             // logger.debug(`Fetching traefik config for exit node: ${currentExitNode}`);
             traefikConfig = await getTraefikConfig(
@@ -528,7 +531,8 @@ 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>();

server/lib/traefik/getTraefikConfig.ts

@@ -1,4 +1,4 @@
-import { db, targetHealthCheck, domains } from "@server/db";
+import { db, targetHealthCheck, domains, aiProviders } from "@server/db";
 import {
     and,
     eq,
@@ -45,7 +45,8 @@ 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 resources with their targets and sites in a single optimized query
     // Start from sites on this exit node, then join to targets and resources
@@ -87,7 +88,7 @@ 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 +210,37 @@ 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
+        .select({
+            resourceId: resources.resourceId,
+            resourceName: resources.name,
+            fullDomain: resources.fullDomain,
+            ssl: resources.ssl,
+            subdomain: resources.subdomain,
+            domainId: resources.domainId,
+            enabled: resources.enabled,
+            domainCertResolver: domains.certResolver,
+            preferWildcardCert: domains.preferWildcardCert
+        })
+        .from(resources)
+        .innerJoin(
+            aiProviders,
+            eq(resources.aiProviderId, 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 +703,90 @@ export async function getTraefikConfig(
             };
         }
     }
+
+    if (aiGatewayUrl) {
+        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 = `${irKey}-service`;
+            const rule = `Host(\`${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 additionalMiddlewares =
+                config.getRawConfig().traefik.additional_middlewares || [];
+            const routerMiddlewares = [
+                badgerMiddlewareName,
+                ...additionalMiddlewares
+            ];
+
+            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 } : {})
+            };
+
+            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.services[serviceName] = {
+                loadBalancer: {
+                    servers: [{ url: `${aiGatewayUrl}/chat/completions` }],
+                    passHostHeader: true
+                }
+            };
+        }
+    }
+
     return config_output;
 }

server/middlewares/index.ts

@@ -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";

server/middlewares/integration/index.ts

@@ -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";

server/middlewares/integration/verifyApiKeyAiModelAccess.ts

@@ -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"
+            )
+        );
+    }
+}

server/middlewares/integration/verifyApiKeyAiProviderAccess.ts

@@ -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"
+            )
+        );
+    }
+}

server/middlewares/integration/verifyApiKeyTargetAccess.ts

@@ -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;
+        const { resourceId, providerId } = target;
-        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`
-                )
-            );
-        }
-
-        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 (apiKey.isRoot) {
-            // Root keys can access any key in any org
+            // Root keys can access any target
             return next();
         }
 
-        if (!resource.orgId) {
+        let orgId: string;
-            return next(
+        if (resourceId) {
-                createHttpError(
+            const [resource] = await db
-                    HttpCode.INTERNAL_SERVER_ERROR,
+                .select()
-                    `Resource with ID ${resourceId} does not have an organization ID`
+                .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,

server/middlewares/verifyAiModelAccess.ts

@@ -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"
+            )
+        );
+    }
+}

server/middlewares/verifyAiProviderAccess.ts

@@ -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"
+            )
+        );
+    }
+}

server/middlewares/verifyTargetAccess.ts

@@ -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(
+        return next(createHttpError(HttpCode.BAD_REQUEST, "Invalid target ID"));
-            createHttpError(HttpCode.BAD_REQUEST, "Invalid organization 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
+        let orgId: string;
-            .select()
-            .from(resources)
-            .where(eq(resources.resourceId, resourceId!))
-            .limit(1);
 
-        if (resource.length === 0) {
+        if (resourceId) {
-            return next(
+            const [resource] = await db
-                createHttpError(
+                .select()
-                    HttpCode.NOT_FOUND,
+                .from(resources)
-                    `Resource with ID ${resourceId} not found`
+                .where(eq(resources.resourceId, resourceId))
-                )
+                .limit(1);
-            );
-        }
 
-        if (!resource[0].orgId) {
+            if (!resource) {
-            return next(
+                return next(
-                createHttpError(
+                    createHttpError(
-                    HttpCode.INTERNAL_SERVER_ERROR,
+                        HttpCode.NOT_FOUND,
-                    `resource with ID ${resourceId} does not have an organization ID`
+                        `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(
+                    and(eq(userOrgs.userId, userId), eq(userOrgs.orgId, orgId))
-                        eq(userOrgs.userId, userId),
-                        eq(userOrgs.orgId, resource[0].orgId)
-                    )
                 );
-            req.userOrg = res[0];
+            req.userOrg = userOrgResult[0];
         }
 
-        if (!req.userOrg) {
+        if (!req.userOrg || req.userOrg.orgId !== orgId) {
-            next(
+            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({
+        if (resourceId) {
-            userId,
+            const resourceAllowed = await canUserAccessResource({
-            resourceId,
+                userId,
-            roleIds: req.userOrgRoleIds ?? []
+                resourceId,
-        });
+                roleIds: req.userOrgRoleIds ?? []
+            });
 
-        if (!resourceAllowed) {
+            if (!resourceAllowed) {
-            return next(
+                return next(
-                createHttpError(
+                    createHttpError(
-                    HttpCode.FORBIDDEN,
+                        HttpCode.FORBIDDEN,
-                    "User does not have access to this resource"
+                        "User does not have access to this resource"
-                )
+                    )
-            );
+                );
+            }
         }
 
-        next();
+        return next();
     } catch (e) {
         return next(
             createHttpError(

server/openApi.ts

@@ -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

server/private/lib/exitNodes/exitNodes.ts

@@ -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.
  *

server/private/lib/traefik/getTraefikConfig.ts

@@ -18,6 +18,7 @@ import {
     domains,
     exitNodes,
     loginPage,
+    SiteResource,
     targetHealthCheck
 } from "@server/db";
 import {
@@ -40,7 +41,8 @@ import {
     siteNetworks,
     siteResources,
     Target,
-    targets
+    targets,
+    aiProviders
 } from "@server/db";
 import {
     sanitize,
@@ -85,7 +87,8 @@ 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 resources with their targets and sites in a single optimized query
     // Start from sites on this exit node, then join to targets and resources
@@ -134,7 +137,7 @@ 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,
@@ -359,7 +362,7 @@ export async function getTraefikConfig(
     let siteResourcesWithFullDomain: {
         siteResourceId: number;
         fullDomain: string | null;
-        mode: "http" | "host" | "cidr" | "ssh";
+        mode: SiteResource["mode"];
     }[] = [];
     if (
         build == "enterprise" &&
@@ -392,6 +395,65 @@ 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
+        .select({
+            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(
+            aiProviders,
+            eq(resources.aiProviderId, aiProviders.providerId)
+        )
+        .leftJoin(domains, eq(domains.domainId, resources.domainId))
+        .where(
+            and(
+                eq(resources.mode, "inference"),
+                eq(resources.enabled, true),
+                eq(aiProviders.enabled, true)
+            )
+        );
+
+    let siteResourcesInference: {
+        siteResourceId: number;
+        alias: string | null;
+        ssl: boolean | null;
+        enabled: boolean | null;
+    }[] = [];
+    if (build == "enterprise") {
+        siteResourcesInference = await db
+            .select({
+                siteResourceId: siteResources.siteResourceId,
+                alias: siteResources.alias,
+                ssl: siteResources.ssl,
+                enabled: siteResources.enabled
+            })
+            .from(siteResources)
+            .innerJoin(
+                aiProviders,
+                eq(siteResources.aiProviderId, aiProviders.providerId)
+            )
+            .where(
+                and(
+                    eq(siteResources.mode, "inference"),
+                    eq(siteResources.enabled, true),
+                    eq(aiProviders.enabled, true),
+                    isNotNull(siteResources.alias)
+                )
+            );
+    }
+
     let validCerts: CertificateResult[] = [];
     if (privateConfig.getRawPrivateConfig().flags.use_pangolin_dns) {
         // create a list of all domains to get certs for
@@ -413,6 +475,17 @@ 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.alias) {
+                domains.add(sr.alias);
+            }
+        }
         // 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)}`);
@@ -1444,6 +1517,199 @@ export async function getTraefikConfig(
         }
     }
 
+    if (aiGatewayUrl) {
+        // 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: ${fullDomain}`
+                    );
+                    continue;
+                }
+            }
+
+            const additionalMiddlewares =
+                config.getRawConfig().traefik.additional_middlewares || [];
+            const routerMiddlewares = [
+                badgerMiddlewareName,
+                ...additionalMiddlewares
+            ];
+
+            if (ir.ssl) {
+                config_output.http.routers[routerName + "-redirect"] = {
+                    entryPoints: [
+                        config.getRawConfig().traefik.http_entrypoint
+                    ],
+                    middlewares: [redirectHttpsMiddlewareName],
+                    service: serviceName,
+                    rule,
+                    priority: 100
+                };
+            }
+
+            config_output.http.routers[routerName] = {
+                entryPoints: [
+                    ir.ssl
+                        ? config.getRawConfig().traefik.https_entrypoint
+                        : config.getRawConfig().traefik.http_entrypoint
+                ],
+                middlewares: routerMiddlewares,
+                service: serviceName,
+                rule,
+                priority: 100,
+                ...(ir.ssl ? { tls } : {})
+            };
+
+            config_output.http.services[serviceName] = {
+                loadBalancer: {
+                    servers: [{ url: `${aiGatewayUrl}/chat/completions` }],
+                    passHostHeader: true
+                }
+            };
+        }
+
+        // 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.alias) continue;
+
+            if (!config_output.http.routers) config_output.http.routers = {};
+            if (!config_output.http.services) config_output.http.services = {};
+
+            const alias = sr.alias;
+            const srKey = `inference-sr${sr.siteResourceId}`;
+            const routerName = `${srKey}-router`;
+            const serviceName = `${srKey}-service`;
+            const rule = `Host(\`${alias}\`)`;
+
+            let tls: any = {};
+            if (!privateConfig.getRawPrivateConfig().flags.use_pangolin_dns) {
+                const domainParts = alias.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 === alias
+                );
+                if (!matchingCert) {
+                    logger.debug(
+                        `No matching certificate found for inference siteResource alias: ${alias}`
+                    );
+                    continue;
+                }
+            }
+
+            const additionalMiddlewares =
+                config.getRawConfig().traefik.additional_middlewares || [];
+
+            if (sr.ssl) {
+                config_output.http.routers[routerName + "-redirect"] = {
+                    entryPoints: [
+                        config.getRawConfig().traefik.http_entrypoint
+                    ],
+                    middlewares: [redirectHttpsMiddlewareName],
+                    service: serviceName,
+                    rule,
+                    priority: 100
+                };
+            }
+
+            config_output.http.routers[routerName] = {
+                entryPoints: [
+                    sr.ssl
+                        ? config.getRawConfig().traefik.https_entrypoint
+                        : config.getRawConfig().traefik.http_entrypoint
+                ],
+                middlewares: additionalMiddlewares,
+                service: serviceName,
+                rule,
+                priority: 100,
+                ...(sr.ssl ? { tls } : {})
+            };
+
+            config_output.http.services[serviceName] = {
+                loadBalancer: {
+                    servers: [{ url: `${aiGatewayUrl}/chat/completions` }],
+                    passHostHeader: true
+                }
+            };
+        }
+    }
+
     if (generateLoginPageRouters) {
         const exitNodeLoginPages = await db
             .select({

server/private/routers/certificates/getBatchedCertificates.ts

@@ -46,7 +46,7 @@ const getCertificateQuerySchema = z.object({
 
 async function query(orgId: string, domainList: string[]) {
     // Try to get CNAME certificates first
-    let existingCertificates = await db
+    const existingCertificates = await db
         .select({
             certId: certificates.certId,
             domain: certificates.domain,
@@ -73,16 +73,19 @@ async function query(orgId: string, domainList: string[]) {
         .where(and(inArray(certificates.domain, domainList)));
 
     // All non resolved domain certificates might be `ns` or `wildcard`,
-    // which means exact domain certificates do not
+    // which means exact domain certificates do not exist
-    const nonAvailableCertificates = existingCertificates
+    const foundDomains = new Set(
-        .filter((cert) => !domainList.includes(cert.domain))
+        existingCertificates.map((cert) => cert.domain)
-        .map((cert) => cert.domain);
+    );
+    const domainsWithMissingCertificates = domainList.filter(
+        (domain) => !foundDomains.has(domain)
+    );
 
-    if (nonAvailableCertificates.length > 0) {
+    if (domainsWithMissingCertificates.length > 0) {
         const domainLevelDownSet = new Set<string>();
         const wildcardDomainSet = new Set<string>();
 
-        for (const domain of nonAvailableCertificates) {
+        for (const domain of domainsWithMissingCertificates) {
             const domainLevelDown = domain.split(".").slice(1).join(".");
             const wildcardPrefixed = `*.${domainLevelDown}`;
             domainLevelDownSet.add(domainLevelDown);
@@ -131,6 +134,7 @@ async function query(orgId: string, domainList: string[]) {
     for (const domain of domainList) {
         const domainLevelDown = domain.split(".").slice(1).join(".");
         const wildcardPrefixed = `*.${domainLevelDown}`;
+
         certificateMap[domain] =
             existingCertificates.find(
                 (cert) =>

server/private/routers/hybrid.ts

@@ -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, {

server/private/routers/re-key/reGenerateSiteSecret.ts

@@ -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,

server/routers/aiGateway/chatCompletions.ts

@@ -0,0 +1,379 @@
+import { Request, Response } from "express";
+import { and, eq } from "drizzle-orm";
+import {
+    AiProvider,
+    aiModels,
+    aiProviders,
+    db,
+    resourceAiModels,
+    resources,
+    siteResourceAiModels,
+    siteResources,
+    users
+} from "@server/db";
+import config from "@server/lib/config";
+import { decrypt } from "@server/lib/crypto";
+import {
+    AiProviderAuthType,
+    AiProviderRoutingMode,
+    AiProviderType,
+    resolveAiProviderConfig
+} from "@server/lib/aiProviderDefaults";
+import { verifyResourceAccessToken } from "@server/auth/verifyResourceAccessToken";
+import {
+    SESSION_COOKIE_NAME,
+    validateSessionToken
+} from "@server/auth/sessions/app";
+import { getUserOrgRoles } from "@server/lib/userOrgRoles";
+import logger from "@server/logger";
+import HttpCode from "@server/types/HttpCode";
+
+type ResolvedTarget = {
+    resourceId: number | null;
+    orgId: string | null;
+    provider: AiProvider;
+    // null = no restriction; every enabled model on the provider is allowed
+    allowedModelIds: number[] | null;
+};
+
+// Fallback for clients that hit the endpoint directly (e.g. an AI tool's
+// "API key" field) instead of going through a browser session - badger
+// forwards whatever Authorization header the client sent untouched in that
+// case. This prefix lets us tell "this bearer value is a Pangolin resource
+// access token" apart from an arbitrary/opaque API key a user might paste
+// in, without guessing based on format alone.
+const USER_TOKEN_PREFIX = "pu_";
+
+export type RequestUser = {
+    userId: string;
+    username: string;
+    email: string | null;
+    name: string | null;
+    role: string | null;
+};
+
+async function buildRequestUser(
+    userId: string,
+    orgId: string | null
+): Promise<RequestUser | null> {
+    const [user] = await db
+        .select()
+        .from(users)
+        .where(eq(users.userId, userId))
+        .limit(1);
+
+    if (!user) {
+        return null;
+    }
+
+    const orgRoles = orgId ? await getUserOrgRoles(user.userId, orgId) : [];
+
+    return {
+        userId: user.userId,
+        username: user.username,
+        email: user.email,
+        name: user.name,
+        role: orgRoles.map((r) => r.roleName).join(", ") || null
+    };
+}
+
+async function resolveRequestUser(
+    req: Request,
+    resourceId: number | null,
+    orgId: string | null
+): Promise<RequestUser | null> {
+    // Public resources behind badger: badger passes the resource session
+    // cookie through to the backend (same mechanism the browser gateway,
+    // e.g. the SSH page, relies on), so we can validate it exactly like
+    // verifySessionUserMiddleware does for the dashboard.
+    const sessionToken = req.cookies?.[SESSION_COOKIE_NAME];
+    if (sessionToken) {
+        const { session, user } = await validateSessionToken(sessionToken);
+        if (session && user) {
+            return buildRequestUser(user.userId, orgId);
+        }
+    }
+
+    // User devices hitting the endpoint directly (no browser session to
+    // forward) fall back to a Pangolin resource access token passed as the
+    // client's "API key".
+    const authHeader = req.headers["authorization"];
+    if (typeof authHeader !== "string" || !resourceId) {
+        return null;
+    }
+
+    const bearer = authHeader.match(/^Bearer\s+(.+)$/i)?.[1];
+    if (!bearer || !bearer.startsWith(USER_TOKEN_PREFIX)) {
+        return null;
+    }
+
+    const [accessTokenId, accessToken] = bearer
+        .slice(USER_TOKEN_PREFIX.length)
+        .split(".");
+    if (!accessTokenId || !accessToken) {
+        return null;
+    }
+
+    const { valid, tokenItem } = await verifyResourceAccessToken({
+        accessToken,
+        accessTokenId,
+        resourceId
+    });
+
+    if (!valid || !tokenItem?.userId) {
+        return null;
+    }
+
+    return buildRequestUser(tokenItem.userId, tokenItem.orgId);
+}
+
+async function resolveTarget(host: string): Promise<ResolvedTarget | null> {
+    const [resourceRow] = await db
+        .select({
+            resourceId: resources.resourceId,
+            orgId: resources.orgId,
+            provider: aiProviders
+        })
+        .from(resources)
+        .innerJoin(
+            aiProviders,
+            eq(resources.aiProviderId, aiProviders.providerId)
+        )
+        .where(
+            and(
+                eq(resources.fullDomain, host),
+                eq(resources.mode, "inference"),
+                eq(resources.enabled, true),
+                eq(aiProviders.enabled, true)
+            )
+        )
+        .limit(1);
+
+    if (resourceRow) {
+        const restrictions = await db
+            .select({ modelId: resourceAiModels.modelId })
+            .from(resourceAiModels)
+            .where(eq(resourceAiModels.resourceId, resourceRow.resourceId));
+
+        return {
+            resourceId: resourceRow.resourceId,
+            orgId: resourceRow.orgId,
+            provider: resourceRow.provider,
+            allowedModelIds: restrictions.length
+                ? restrictions.map((r) => r.modelId)
+                : null
+        };
+    }
+
+    const [siteResourceRow] = await db
+        .select({
+            siteResourceId: siteResources.siteResourceId,
+            orgId: siteResources.orgId,
+            provider: aiProviders
+        })
+        .from(siteResources)
+        .innerJoin(
+            aiProviders,
+            eq(siteResources.aiProviderId, aiProviders.providerId)
+        )
+        .where(
+            and(
+                eq(siteResources.alias, host),
+                eq(siteResources.mode, "inference"),
+                eq(siteResources.enabled, true),
+                eq(aiProviders.enabled, true)
+            )
+        )
+        .limit(1);
+
+    if (siteResourceRow) {
+        const restrictions = await db
+            .select({ modelId: siteResourceAiModels.modelId })
+            .from(siteResourceAiModels)
+            .where(
+                eq(
+                    siteResourceAiModels.siteResourceId,
+                    siteResourceRow.siteResourceId
+                )
+            );
+
+        return {
+            // siteResources have no per-user auth/policy stack today (see
+            // the routing comment in getTraefikConfig.ts), so there's no
+            // resource access token scope to validate a user token against.
+            resourceId: null,
+            orgId: siteResourceRow.orgId,
+            provider: siteResourceRow.provider,
+            allowedModelIds: restrictions.length
+                ? restrictions.map((r) => r.modelId)
+                : null
+        };
+    }
+
+    return null;
+}
+
+// Generic OpenAI-wire-compatible passthrough. Anthropic's native API uses a
+// different path/schema; everything else here is OpenAI-compatible today.
+function getCompletionsPath(type: AiProviderType): string {
+    if (type === "anthropic") {
+        return "/v1/messages";
+    }
+    return "/chat/completions";
+}
+
+export async function chatCompletions(req: Request, res: Response): Promise<any> {
+    try {
+        const host = (req.headers.host || "").split(":")[0];
+        if (!host) {
+            return res
+                .status(HttpCode.BAD_REQUEST)
+                .json({ error: { message: "Missing Host header" } });
+        }
+
+        const target = await resolveTarget(host);
+        if (!target) {
+            return res.status(HttpCode.NOT_FOUND).json({
+                error: {
+                    message: "No inference resource found for this host"
+                }
+            });
+        }
+
+        const { provider, allowedModelIds, resourceId, orgId } = target;
+
+        // Best-effort identity resolution - not yet enforced, but lets us
+        // start making per-user access decisions (e.g. model/role-based
+        // restrictions) without another round of plumbing later.
+        const requestUser = await resolveRequestUser(req, resourceId, orgId);
+        if (requestUser) {
+            logger.debug(
+                `AI gateway request from user ${requestUser.userId} (${requestUser.username})`
+            );
+        }
+
+        const requestedModel =
+            typeof req.body?.model === "string" ? req.body.model : undefined;
+
+        if (allowedModelIds) {
+            if (!requestedModel) {
+                return res.status(HttpCode.FORBIDDEN).json({
+                    error: {
+                        message:
+                            "This resource restricts access to specific models; a model must be specified"
+                    }
+                });
+            }
+
+            const [matchedModel] = await db
+                .select({ modelId: aiModels.modelId })
+                .from(aiModels)
+                .where(
+                    and(
+                        eq(aiModels.providerId, provider.providerId),
+                        eq(aiModels.modelKey, requestedModel)
+                    )
+                )
+                .limit(1);
+
+            if (
+                !matchedModel ||
+                !allowedModelIds.includes(matchedModel.modelId)
+            ) {
+                return res.status(HttpCode.FORBIDDEN).json({
+                    error: {
+                        message: `Model "${requestedModel}" is not permitted on this resource`
+                    }
+                });
+            }
+        }
+
+        if (!provider.apiKey) {
+            return res.status(HttpCode.INTERNAL_SERVER_ERROR).json({
+                error: { message: "AI provider has no API key configured" }
+            });
+        }
+
+        const secret = config.getRawConfig().server.secret!;
+        const apiKey = decrypt(provider.apiKey, secret);
+
+        const { upstreamUrl, authType } = resolveAiProviderConfig({
+            type: provider.type as AiProviderType,
+            upstreamUrl: provider.upstreamUrl,
+            authType: provider.authType as AiProviderAuthType | null,
+            routingMode: provider.routingMode as AiProviderRoutingMode | null
+        });
+
+        if (!upstreamUrl) {
+            return res.status(HttpCode.INTERNAL_SERVER_ERROR).json({
+                error: {
+                    message: "AI provider has no upstream URL configured"
+                }
+            });
+        }
+
+        const targetUrl = `${upstreamUrl.replace(/\/$/, "")}${getCompletionsPath(
+            provider.type as AiProviderType
+        )}`;
+
+        const headers: Record<string, string> = {
+            "Content-Type": "application/json"
+        };
+        if (authType === "bearer") {
+            headers["Authorization"] = `Bearer ${apiKey}`;
+        }
+
+        // No dedicated per-request TLS agent is wired up (no extra deps for
+        // this v1 gateway) - toggle the process-wide Node TLS check instead.
+        // Known limitation: this is not safe under concurrent requests mixing
+        // skipTlsVerification providers with strict ones.
+        const restoreTlsReject = process.env.NODE_TLS_REJECT_UNAUTHORIZED;
+        if (provider.skipTlsVerification) {
+            process.env.NODE_TLS_REJECT_UNAUTHORIZED = "0";
+        }
+
+        let upstreamRes: globalThis.Response;
+        try {
+            upstreamRes = await fetch(targetUrl, {
+                method: "POST",
+                headers,
+                body: JSON.stringify(req.body)
+            });
+        } finally {
+            if (provider.skipTlsVerification) {
+                if (restoreTlsReject === undefined) {
+                    delete process.env.NODE_TLS_REJECT_UNAUTHORIZED;
+                } else {
+                    process.env.NODE_TLS_REJECT_UNAUTHORIZED = restoreTlsReject;
+                }
+            }
+        }
+
+        const contentType = upstreamRes.headers.get("content-type") || "";
+        const isStream =
+            req.body?.stream === true ||
+            contentType.includes("text/event-stream");
+
+        res.status(upstreamRes.status);
+        res.setHeader("Content-Type", contentType || "application/json");
+
+        if (isStream && upstreamRes.body) {
+            res.flushHeaders();
+            const reader = upstreamRes.body.getReader();
+            while (true) {
+                const { done, value } = await reader.read();
+                if (done) break;
+                res.write(value);
+            }
+            return res.end();
+        }
+
+        const text = await upstreamRes.text();
+        return res.send(text);
+    } catch (error) {
+        logger.error(error);
+        return res.status(HttpCode.INTERNAL_SERVER_ERROR).json({
+            error: { message: "Failed to proxy inference request" }
+        });
+    }
+}

server/routers/aiGateway/index.ts

@@ -0,0 +1 @@
+export * from "./chatCompletions";

server/routers/aiProvider/createAiModel.ts

@@ -0,0 +1,151 @@
+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 {
+    aiBudgetUnitSchema,
+    refineBudgetFields
+} from "@server/routers/aiProvider/validation";
+
+const paramsSchema = z.strictObject({
+    providerId: z.coerce.number().int().positive()
+});
+
+const bodySchema = z
+    .strictObject({
+        modelKey: z.string().nonempty(),
+        name: z.string().nonempty(),
+        budgetAmount: z.number().positive().optional().nullable(),
+        budgetUnit: aiBudgetUnitSchema.optional().nullable(),
+        enabled: z.boolean().optional()
+    })
+    .superRefine((data, ctx) => {
+        refineBudgetFields(data, ctx);
+    });
+
+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, budgetAmount, budgetUnit, enabled } =
+            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,
+                budgetAmount: budgetAmount ?? null,
+                budgetUnit: budgetUnit ?? null,
+                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")
+        );
+    }
+}

server/routers/aiProvider/createAiProvider.ts

@@ -0,0 +1,150 @@
+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 type { CreateOrEditAiProviderResponse } from "@server/routers/aiProvider/types";
+import { toPublicAiProvider } from "@server/routers/aiProvider/types";
+import {
+    aiAuthTypeSchema,
+    aiBudgetUnitSchema,
+    aiProviderTypeSchema,
+    aiRoutingModeSchema,
+    refineBudgetFields,
+    refineProviderUpstreamFields
+} from "@server/routers/aiProvider/validation";
+
+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().nullable(),
+        routingMode: aiRoutingModeSchema.optional(),
+        skipTlsVerification: z.boolean().optional(),
+        budgetAmount: z.number().positive().optional().nullable(),
+        budgetUnit: aiBudgetUnitSchema.optional().nullable(),
+        enabled: z.boolean().optional()
+    })
+    .superRefine((data, ctx) => {
+        refineProviderUpstreamFields(data, ctx);
+        refineBudgetFields(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,
+            skipTlsVerification,
+            budgetAmount,
+            budgetUnit,
+            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 resolvedRoutingMode =
+            type === "custom" ? (routingMode ?? "url") : "url";
+
+        const [provider] = await db
+            .insert(aiProviders)
+            .values({
+                orgId,
+                name,
+                type,
+                upstreamUrl:
+                    resolvedRoutingMode === "target"
+                        ? null
+                        : (upstreamUrl ?? null),
+                apiKey: encryptedApiKey,
+                apiKeyLastChars,
+                authType: authType ?? null,
+                routingMode: resolvedRoutingMode,
+                skipTlsVerification: skipTlsVerification ?? false,
+                budgetAmount: budgetAmount ?? null,
+                budgetUnit: budgetUnit ?? null,
+                enabled: enabled ?? true,
+                createdAt: now,
+                updatedAt: now
+            })
+            .returning();
+
+        return response<CreateOrEditAiProviderResponse>(res, {
+            data: { provider: toPublicAiProvider(provider) },
+            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")
+        );
+    }
+}

server/routers/aiProvider/deleteAiModel.ts

@@ -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")
+        );
+    }
+}

server/routers/aiProvider/deleteAiProvider.ts

@@ -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")
+        );
+    }
+}

server/routers/aiProvider/getAiModel.ts

@@ -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")
+        );
+    }
+}

server/routers/aiProvider/getAiProvider.ts

@@ -0,0 +1,82 @@
+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) },
+            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")
+        );
+    }
+}

server/routers/aiProvider/index.ts

@@ -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";

server/routers/aiProvider/listAiModels.ts

@@ -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")
+        );
+    }
+}

server/routers/aiProvider/listAiProviders.ts

@@ -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(toPublicAiProvider),
+                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")
+        );
+    }
+}

server/routers/aiProvider/types.ts

@@ -0,0 +1,53 @@
+import type { AiModel, AiProvider } from "@server/db";
+import type { PaginatedResponse } from "@server/types/Pagination";
+import {
+    resolveAiProviderConfig,
+    type AiProviderAuthType,
+    type AiProviderRoutingMode,
+    type AiProviderType
+} from "@server/lib/aiProviderDefaults";
+
+export type AiProviderPublic = Omit<AiProvider, "apiKey"> & {
+    effectiveUpstreamUrl: string | null;
+    effectiveAuthType: AiProviderAuthType | null;
+};
+
+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): AiProviderPublic {
+    const { apiKey: _apiKey, ...rest } = provider;
+    const resolved = resolveAiProviderConfig({
+        type: provider.type as AiProviderType,
+        upstreamUrl: provider.upstreamUrl,
+        authType: provider.authType as AiProviderAuthType | null,
+        routingMode: provider.routingMode as AiProviderRoutingMode | null
+    });
+
+    return {
+        ...rest,
+        effectiveUpstreamUrl: resolved.upstreamUrl,
+        effectiveAuthType: resolved.authType
+    };
+}

server/routers/aiProvider/updateAiModel.ts

@@ -0,0 +1,167 @@
+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 {
+    aiBudgetUnitSchema,
+    refineBudgetFields
+} from "@server/routers/aiProvider/validation";
+
+const paramsSchema = z.strictObject({
+    modelId: z.coerce.number().int().positive()
+});
+
+const bodySchema = z
+    .strictObject({
+        modelKey: z.string().nonempty().optional(),
+        name: z.string().nonempty().optional(),
+        budgetAmount: z.number().positive().optional().nullable(),
+        budgetUnit: aiBudgetUnitSchema.optional().nullable(),
+        enabled: z.boolean().optional()
+    })
+    .superRefine((data, ctx) => {
+        refineBudgetFields(data, ctx);
+    });
+
+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.budgetAmount !== undefined) {
+            updateData.budgetAmount = body.budgetAmount;
+        }
+        if (body.budgetUnit !== undefined) {
+            updateData.budgetUnit = body.budgetUnit;
+        }
+        if (body.enabled !== undefined) {
+            updateData.enabled = body.enabled;
+        }
+
+        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")
+        );
+    }
+}

server/routers/aiProvider/updateAiProvider.ts

@@ -0,0 +1,208 @@
+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,
+    aiBudgetUnitSchema,
+    aiProviderTypeSchema,
+    aiRoutingModeSchema,
+    refineBudgetFields,
+    refineProviderUpstreamFields
+} from "@server/routers/aiProvider/validation";
+import type {
+    AiProviderRoutingMode,
+    AiProviderType
+} from "@server/lib/aiProviderDefaults";
+
+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().nullable(),
+        routingMode: aiRoutingModeSchema.optional(),
+        skipTlsVerification: z.boolean().optional(),
+        budgetAmount: z.number().positive().optional().nullable(),
+        budgetUnit: aiBudgetUnitSchema.optional().nullable(),
+        enabled: z.boolean().optional()
+    })
+    .superRefine((data, ctx) => {
+        refineBudgetFields(data, ctx);
+    });
+
+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 ${providerId} not found`
+                )
+            );
+        }
+
+        const providerType = existing.type as AiProviderType;
+        const nextRoutingMode: AiProviderRoutingMode =
+            providerType === "custom"
+                ? ((body.routingMode ??
+                      existing.routingMode) as AiProviderRoutingMode)
+                : "url";
+        const nextUpstreamUrl =
+            body.upstreamUrl !== undefined
+                ? body.upstreamUrl
+                : existing.upstreamUrl;
+        const nextAuthType =
+            body.authType !== undefined ? body.authType : existing.authType;
+
+        const validation = z
+            .object({
+                type: aiProviderTypeSchema,
+                upstreamUrl: z.string().nullable().optional(),
+                authType: aiAuthTypeSchema.nullable().optional(),
+                routingMode: aiRoutingModeSchema.optional()
+            })
+            .superRefine((data, ctx) => refineProviderUpstreamFields(data, ctx))
+            .safeParse({
+                type: providerType,
+                upstreamUrl: nextUpstreamUrl,
+                authType: nextAuthType,
+                routingMode: nextRoutingMode
+            });
+
+        if (!validation.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(validation.error).toString()
+                )
+            );
+        }
+
+        const updateData: Partial<typeof aiProviders.$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 (body.budgetAmount !== undefined) {
+            updateData.budgetAmount = body.budgetAmount;
+        }
+        if (body.budgetUnit !== undefined) {
+            updateData.budgetUnit = body.budgetUnit;
+        }
+        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.apiKey !== undefined) {
+            const key = config.getRawConfig().server.secret!;
+            updateData.apiKey = encrypt(body.apiKey, key);
+            updateData.apiKeyLastChars = body.apiKey.slice(-4);
+        }
+
+        const [provider] = await db
+            .update(aiProviders)
+            .set(updateData)
+            .where(eq(aiProviders.providerId, providerId))
+            .returning();
+
+        return response<CreateOrEditAiProviderResponse>(res, {
+            data: { provider: toPublicAiProvider(provider) },
+            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")
+        );
+    }
+}

server/routers/aiProvider/validation.ts

@@ -0,0 +1,85 @@
+import { z } from "zod";
+import {
+    providerRequiresUpstreamUrl,
+    type AiBudgetUnit,
+    type AiProviderRoutingMode,
+    type AiProviderType
+} from "@server/lib/aiProviderDefaults";
+
+export const aiProviderTypeSchema = z.enum([
+    "openai",
+    "anthropic",
+    "googleGemini",
+    "vertexAi",
+    "bedrock",
+    "microsoftFoundry",
+    "openRouter",
+    "vercelAiGateway",
+    "custom"
+]);
+
+export const aiBudgetUnitSchema = z.enum(["usd", "tokens"]);
+
+export const aiAuthTypeSchema = z.enum(["bearer"]);
+
+export const aiRoutingModeSchema = z.enum(["url", "target"]);
+
+export function refineBudgetFields(
+    data: {
+        budgetAmount?: number | null;
+        budgetUnit?: AiBudgetUnit | null;
+    },
+    ctx: z.RefinementCtx
+) {
+    const hasAmount =
+        data.budgetAmount !== undefined && data.budgetAmount !== null;
+    const hasUnit = data.budgetUnit !== undefined && data.budgetUnit !== null;
+
+    if (hasAmount !== hasUnit) {
+        ctx.addIssue({
+            code: "custom",
+            message:
+                "budgetAmount and budgetUnit must both be set or both omitted",
+            path: hasAmount ? ["budgetUnit"] : ["budgetAmount"]
+        });
+    }
+}
+
+export function refineProviderUpstreamFields(
+    data: {
+        type: AiProviderType;
+        upstreamUrl?: string | null;
+        authType?: "bearer" | null;
+        routingMode?: AiProviderRoutingMode | 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" && routingMode === "url" && !data.authType) {
+        ctx.addIssue({
+            code: "custom",
+            message: "authType is required for custom providers",
+            path: ["authType"]
+        });
+    }
+}

server/routers/badger/verifySession.ts

@@ -219,7 +219,9 @@ 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);

server/routers/client/createClient.ts

@@ -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,

server/routers/client/createUserClient.ts

@@ -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,

server/routers/client/listUserDevices.ts

@@ -67,12 +67,12 @@ const listUserDevicesSchema = z.strictObject({
         }),
     query: z.string().optional(),
     sort_by: z
-        .enum(["megabytesIn", "megabytesOut"])
+        .enum(["megabytesIn", "megabytesOut", "firstSeen", "lastSeen"])
         .optional()
         .catch(undefined)
         .openapi({
             type: "string",
-            enum: ["megabytesIn", "megabytesOut"],
+            enum: ["megabytesIn", "megabytesOut", "firstSeen", "lastSeen"],
             description: "Field to sort by"
         }),
     order: z
@@ -183,7 +183,9 @@ function queryUserDevicesBase() {
             fingerprintArch: currentFingerprint.arch,
             fingerprintSerialNumber: currentFingerprint.serialNumber,
             fingerprintUsername: currentFingerprint.username,
-            fingerprintHostname: currentFingerprint.hostname
+            fingerprintHostname: currentFingerprint.hostname,
+            firstSeen: currentFingerprint.firstSeen,
+            lastSeen: currentFingerprint.lastSeen
         })
         .from(clients)
         .leftJoin(orgs, eq(clients.orgId, orgs.orgId))
@@ -389,14 +391,23 @@ export async function listUserDevices(
 
         const countQuery = db.$count(baseQuery.as("filtered_clients"));
 
+        const sortColumn =
+            sort_by === "firstSeen"
+                ? currentFingerprint.firstSeen
+                : sort_by === "lastSeen"
+                  ? currentFingerprint.lastSeen
+                  : sort_by
+                    ? clients[sort_by]
+                    : undefined;
+
         const listDevicesQuery = baseQuery
             .limit(pageSize)
             .offset(pageSize * (page - 1))
             .orderBy(
-                sort_by
+                sortColumn
                     ? order === "asc"
-                        ? asc(clients[sort_by])
+                        ? asc(sortColumn)
-                        : desc(clients[sort_by])
+                        : desc(sortColumn)
                     : asc(clients.clientId)
             );
 

server/routers/external.ts

@@ -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 +407,13 @@ authenticated.get(
     siteResource.listSiteResourceClients
 );
 
+authenticated.get(
+    "/site-resource/:siteResourceId/ai-models",
+    verifySiteResourceAccess,
+    verifyUserHasAction(ActionsEnum.listResourceAiModels),
+    siteResource.listSiteResourceAiModels
+);
+
 authenticated.post(
     "/site-resource/:siteResourceId/roles",
     verifySiteResourceAccess,
@@ -414,6 +424,14 @@ 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/users",
     verifySiteResourceAccess,
@@ -648,6 +666,13 @@ authenticated.get(
     resource.listResourceUsers
 );
 
+authenticated.get(
+    "/resource/:resourceId/ai-models",
+    verifyResourceAccess,
+    verifyUserHasAction(ActionsEnum.listResourceAiModels),
+    resource.listResourceAiModels
+);
+
 authenticated.get(
     "/resource/:resourceId",
     verifyResourceAccess,
@@ -851,6 +876,14 @@ authenticated.post(
     resource.setResourceUsers
 );
 
+authenticated.post(
+    "/resource/:resourceId/ai-models",
+    verifyResourceAccess,
+    verifyUserHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    resource.setResourceAiModels
+);
+
 authenticated.put(
     "/resource-policy/:resourcePolicyId/access-control",
     verifyResourcePolicyAccess,
@@ -1366,6 +1399,99 @@ 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,

server/routers/gerbil/getAllRelays.ts

@@ -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
                     });
                 }

server/routers/gerbil/getConfig.ts

@@ -89,7 +89,7 @@ export async function generateGerbilConfig(exitNode: ExitNode) {
             and(
                 eq(sites.exitNodeId, exitNode.exitNodeId),
                 isNotNull(sites.pubKey),
-                isNotNull(sites.subnet)
+                isNotNull(sites.exitNodeSubnet)
             )
         );
 
@@ -103,7 +103,7 @@ export async function generateGerbilConfig(exitNode: ExitNode) {
             } else if (site.type === "newt") {
                 return {
                     publicKey: site.pubKey,
-                    allowedIps: [site.subnet!]
+                    allowedIps: [site.exitNodeSubnet!]
                 };
             }
             return {

server/routers/gerbil/updateHolePunch.ts

@@ -186,7 +186,7 @@ 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,

server/routers/integration.ts

@@ -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 +246,16 @@ authenticated.get(
     siteResource.listSiteResourceClients
 );
 
+authenticated.get(
+    [
+        "/site-resource/:siteResourceId/ai-models",
+        "/private-resource/:siteResourceId/ai-models"
+    ],
+    verifyApiKeySiteResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.listResourceAiModels),
+    siteResource.listSiteResourceAiModels
+);
+
 authenticated.post(
     [
         "/site-resource/:siteResourceId/roles",
@@ -295,6 +308,39 @@ 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/users/add",
@@ -507,6 +553,16 @@ 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", "/public-resource/:resourceId"],
     verifyApiKeyResourceAccess,
@@ -708,6 +764,17 @@ 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/users", "/public-resource/:resourceId/users"],
     verifyApiKeyResourceAccess,
@@ -900,6 +967,28 @@ 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/users/add",
@@ -1366,3 +1455,95 @@ 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
+);

server/routers/internal.ts

@@ -3,6 +3,7 @@ import * as gerbil from "@server/routers/gerbil";
 import * as traefik from "@server/routers/traefik";
 import * as resource from "./resource";
 import * as badger from "./badger";
+import * as aiGateway from "@server/routers/aiGateway";
 import * as auth from "@server/routers/auth";
 import * as supporterKey from "@server/routers/supporterKey";
 import * as idp from "@server/routers/idp";
@@ -63,3 +64,10 @@ internalRouter.use("/badger", badgerRouter);
 badgerRouter.post("/verify-session", badger.verifyResourceSession);
 
 badgerRouter.post("/exchange-session", badger.exchangeSession);
+
+// AI inference gateway - minimal chat-completions proxy for inference-mode
+// resources/siteResources
+internalRouter.post(
+    "/ai-gateway/chat/completions",
+    aiGateway.chatCompletions
+);

server/routers/launcher/formatLauncherAccess.ts

@@ -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;

server/routers/newt/buildConfiguration.ts

@@ -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))
             )
         );
 

server/routers/newt/handleNewtExitNodesRequestMessage.ts

@@ -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)
+            const weight = await calculateExitNodeWeight(
-            // higher = more desirable
+                node.exitNodeId,
-            // like saying, this node has x% of its capacity left
+                node.maxConnections
+            );
 
-            let weight = 1;
+            if (weight === null) {
-            const maxConnections = node.maxConnections;
+                return null;
-            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;
             }
 
             return {

server/routers/newt/handleNewtGetConfigMessage.ts

@@ -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: {

server/routers/newt/handleNewtRegisterMessage.ts

@@ -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
-    );
-}

server/routers/newt/index.ts

@@ -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";

server/routers/olm/buildConfiguration.ts

@@ -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))
+        .leftJoin(networks, eq(siteResources.networkId, networks.networkId))
-        .innerJoin(siteNetworks, eq(networks.networkId, siteNetworks.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.alias)
+        .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
+    };
 }

server/routers/olm/handleOlmExitNodesRequestMessage.ts

@@ -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
+    };
+};

server/routers/olm/handleOlmRegisterMessage.ts

@@ -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 +22,12 @@ 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";
 
 const HOLEPUNCH_STALE_CHAIN_THRESHOLD = 18;
 const HOLEPUNCH_STALE_CHAIN_TTL_SECONDS = 1800;
@@ -49,11 +55,20 @@ 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 +299,64 @@ 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 +366,9 @@ 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));
 
@@ -376,14 +450,28 @@ 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(
+    const { siteConfigurations, exitNodeAliases } =
-        client,
+        await buildSiteConfigurationForOlmClient(
-        publicKey,
+            client,
-        relay,
+            publicKey,
-        jitMode
+            relay,
+            jitMode
+        );
+
+    logger.info(
+        `+++++++++++++++++++++++++++++++ ExitNode Aliases: ${exitNodeAliases}`
     );
 
     // Return connect message with all site configurations
@@ -394,6 +482,17 @@ 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]
+                          }
+                        : undefined,
                 chainId: chainId
             }
         },

server/routers/olm/index.ts

@@ -15,3 +15,4 @@ export * from "./handleOlmServerInitAddPeerHandshake";
 export * from "./offlineChecker";
 export * from "./handleOlmUnLocalMessage";
 export * from "./handleOlmLocalMessage";
+export * from "./handleOlmExitNodesRequestMessage";

server/routers/resource/addAiModelToResource.ts

@@ -0,0 +1,161 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { db, resources, resourceAiModels, aiModels } 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";
+
+const addAiModelToResourceBodySchema = z.strictObject({
+    modelId: z.int().positive()
+});
+
+const addAiModelToResourceParamsSchema = z.strictObject({
+    resourceId: z.coerce.number().int().positive()
+});
+
+registry.registerPath({
+    method: "post",
+    path: "/resource/{resourceId}/ai-models/add",
+    description:
+        "Add a single AI model to a resource's model restriction allow-list.",
+    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 } = 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")
+            );
+        }
+
+        if (!resource.aiProviderId) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    "Resource has no AI provider linked"
+                )
+            );
+        }
+
+        const [model] = await db
+            .select()
+            .from(aiModels)
+            .where(
+                and(
+                    eq(aiModels.modelId, modelId),
+                    eq(aiModels.providerId, resource.aiProviderId)
+                )
+            )
+            .limit(1);
+
+        if (!model) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    "Model not found or does not belong to this resource's AI provider"
+                )
+            );
+        }
+
+        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 });
+
+        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")
+        );
+    }
+}

server/routers/resource/createResource.ts

@@ -44,11 +44,11 @@ const createResourceParamsSchema = z.strictObject({
 });
 
 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 +90,22 @@ 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
+        aiProviderId: z
+            .number()
+            .int()
+            .positive()
+            .optional()
+            .describe(
+                "For inference-mode resources: the AI provider this resource proxies chat completions to."
+            )
     })
     .refine(
         (data) => {
@@ -365,7 +376,8 @@ async function createHttpResource(
         mode,
         authDaemonPort,
         authDaemonMode,
-        pamMode
+        pamMode,
+        aiProviderId
     } = parsedBody.data;
     const subdomain = parsedBody.data.subdomain;
     const stickySession = parsedBody.data.stickySession;
@@ -552,7 +564,8 @@ async function createHttpResource(
                 postAuthPath: postAuthPath,
                 wildcard,
                 health: "unknown",
-                defaultResourcePolicyId: defaultPolicy.resourcePolicyId
+                defaultResourcePolicyId: defaultPolicy.resourcePolicyId,
+                aiProviderId: aiProviderId ?? null
             })
             .returning();
 

server/routers/resource/index.ts

@@ -35,3 +35,7 @@ export * from "./removeEmailFromResourceWhitelist";
 export * from "./getStatusHistory";
 export * from "./getBatchedStatusHistory";
 export * from "./getResourcePolicies";
+export * from "./listResourceAiModels";
+export * from "./setResourceAiModels";
+export * from "./addAiModelToResource";
+export * from "./removeAiModelFromResource";

server/routers/resource/listResourceAiModels.ts

@@ -0,0 +1,107 @@
+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,
+            enabled: aiModels.enabled
+        })
+        .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 AI models a resource is restricted to. An empty list means the resource is not restricted and every enabled model on its linked AI provider is allowed.",
+    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")
+        );
+    }
+}

server/routers/resource/listResources.ts

@@ -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;

server/routers/resource/removeAiModelFromResource.ts

@@ -0,0 +1,141 @@
+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";
+
+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 AI model from a resource's model restriction allow-list.",
+    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 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")
+        );
+    }
+}

server/routers/resource/setResourceAiModels.ts

@@ -0,0 +1,155 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { db, resources, resourceAiModels, aiModels } from "@server/db";
+import { eq, and, inArray } 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 setResourceAiModelsBodySchema = z.strictObject({
+    modelIds: z.array(z.int().positive())
+});
+
+const setResourceAiModelsParamsSchema = z.strictObject({
+    resourceId: z.coerce.number().int().positive()
+});
+
+registry.registerPath({
+    method: "post",
+    path: "/resource/{resourceId}/ai-models",
+    description:
+        "Set the AI models a resource is restricted to. This replaces all existing restrictions. Pass an empty array to remove the restriction (allow every enabled model on the linked provider).",
+    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 { modelIds } = 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")
+            );
+        }
+
+        if (modelIds.length > 0) {
+            if (!resource.aiProviderId) {
+                return next(
+                    createHttpError(
+                        HttpCode.BAD_REQUEST,
+                        "Resource has no AI provider linked"
+                    )
+                );
+            }
+
+            const validModels = await db
+                .select({ modelId: aiModels.modelId })
+                .from(aiModels)
+                .where(
+                    and(
+                        inArray(aiModels.modelId, modelIds),
+                        eq(aiModels.providerId, resource.aiProviderId)
+                    )
+                );
+
+            if (validModels.length !== new Set(modelIds).size) {
+                return next(
+                    createHttpError(
+                        HttpCode.BAD_REQUEST,
+                        "One or more model IDs do not exist or do not belong to this resource's AI provider"
+                    )
+                );
+            }
+        }
+
+        await db.transaction(async (trx) => {
+            await trx
+                .delete(resourceAiModels)
+                .where(eq(resourceAiModels.resourceId, resourceId));
+
+            if (modelIds.length > 0) {
+                await trx
+                    .insert(resourceAiModels)
+                    .values(
+                        modelIds.map((modelId) => ({ resourceId, modelId }))
+                    );
+            }
+        });
+
+        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")
+        );
+    }
+}

server/routers/resource/updateResource.ts

@@ -120,6 +120,15 @@ const updateHttpResourceBodySchema = z
             .optional()
             .describe(
                 "ID of the resource policy to apply to this resource. Set to null to remove the resource policy and fall back to the inline policy settings."
+            ),
+        aiProviderId: z
+            .number()
+            .int()
+            .positive()
+            .nullable()
+            .optional()
+            .describe(
+                "For inference-mode resources: the AI provider this resource proxies chat completions to. Set to null to unlink."
             )
     })
     .refine((data) => Object.keys(data).length > 0, {

server/routers/site/createSite.ts

@@ -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();

server/routers/site/listSites.ts

@@ -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,

server/routers/site/pickSiteDefaults.ts

@@ -43,7 +43,6 @@ const PickSiteDefaultsResponseDataSchema = z.object({
     clientAddress: z.string().optional()
 });
 
-
 registry.registerPath({
     method: "get",
     path: "/org/{orgId}/pick-site-defaults",
@@ -60,7 +59,9 @@ registry.registerPath({
             description: "Successful response",
             content: {
                 "application/json": {
-                    schema: createApiResponseSchema(PickSiteDefaultsResponseDataSchema)
+                    schema: createApiResponseSchema(
+                        PickSiteDefaultsResponseDataSchema
+                    )
                 }
             }
         }
@@ -108,7 +109,7 @@ export async function pickSiteDefaults(
         // list all of the sites on that exit node
         const sitesQuery = await db
             .select({
-                subnet: sites.subnet
+                subnet: sites.exitNodeSubnet
             })
             .from(sites)
             .where(eq(sites.exitNodeId, randomExitNode.exitNodeId));

server/routers/siteResource/addAiModelToSiteResource.ts

@@ -0,0 +1,170 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import {
+    db,
+    siteResources,
+    siteResourceAiModels,
+    aiModels
+} 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";
+
+const addAiModelToSiteResourceBodySchema = z.strictObject({
+    modelId: z.int().positive()
+});
+
+const addAiModelToSiteResourceParamsSchema = z.strictObject({
+    siteResourceId: z.coerce.number().int().positive()
+});
+
+registry.registerPath({
+    method: "post",
+    path: "/site-resource/{siteResourceId}/ai-models/add",
+    description:
+        "Add a single AI model to a site resource's model restriction allow-list.",
+    tags: [OpenAPITags.PrivateResource],
+    request: {
+        params: addAiModelToSiteResourceParamsSchema,
+        body: {
+            content: {
+                "application/json": {
+                    schema: addAiModelToSiteResourceBodySchema
+                }
+            }
+        }
+    },
+    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 addAiModelToSiteResource(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedBody = addAiModelToSiteResourceBodySchema.safeParse(
+            req.body
+        );
+        if (!parsedBody.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedBody.error).toString()
+                )
+            );
+        }
+
+        const { modelId } = parsedBody.data;
+
+        const parsedParams = addAiModelToSiteResourceParamsSchema.safeParse(
+            req.params
+        );
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { siteResourceId } = parsedParams.data;
+
+        const [siteResource] = await db
+            .select()
+            .from(siteResources)
+            .where(eq(siteResources.siteResourceId, siteResourceId))
+            .limit(1);
+
+        if (!siteResource) {
+            return next(
+                createHttpError(HttpCode.NOT_FOUND, "Site resource not found")
+            );
+        }
+
+        if (!siteResource.aiProviderId) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    "Site resource has no AI provider linked"
+                )
+            );
+        }
+
+        const [model] = await db
+            .select()
+            .from(aiModels)
+            .where(
+                and(
+                    eq(aiModels.modelId, modelId),
+                    eq(aiModels.providerId, siteResource.aiProviderId)
+                )
+            )
+            .limit(1);
+
+        if (!model) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    "Model not found or does not belong to this site resource's AI provider"
+                )
+            );
+        }
+
+        const existingEntry = await db
+            .select()
+            .from(siteResourceAiModels)
+            .where(
+                and(
+                    eq(siteResourceAiModels.siteResourceId, siteResourceId),
+                    eq(siteResourceAiModels.modelId, modelId)
+                )
+            );
+
+        if (existingEntry.length > 0) {
+            return next(
+                createHttpError(
+                    HttpCode.CONFLICT,
+                    "Model already assigned to site resource"
+                )
+            );
+        }
+
+        await db
+            .insert(siteResourceAiModels)
+            .values({ siteResourceId, modelId });
+
+        return response(res, {
+            data: {},
+            success: true,
+            error: false,
+            message: "Model added to site resource successfully",
+            status: HttpCode.CREATED
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}

server/routers/siteResource/createSiteResource.ts

@@ -29,7 +29,7 @@ import response from "@server/lib/response";
 import logger from "@server/logger";
 import { OpenAPITags, registry } from "@server/openApi";
 import HttpCode from "@server/types/HttpCode";
-import { and, eq, inArray } from "drizzle-orm";
+import { and, eq, inArray, ne } from "drizzle-orm";
 import { NextFunction, Request, Response } from "express";
 import createHttpError from "http-errors";
 import { z } from "zod";
@@ -49,7 +49,7 @@ const createSiteResourceSchema = z
         name: z.string().min(1).max(255),
         niceId: z.string().optional(),
         // protocol: z.enum(["tcp", "udp"]).optional(),
-        mode: z.enum(["host", "cidr", "http", "ssh"]),
+        mode: z.enum(["host", "cidr", "http", "ssh", "inference"]),
         ssl: z.boolean().optional(), // only used for http mode
         scheme: z.enum(["http", "https"]).optional(),
         siteIds: z.array(z.int()).optional(),
@@ -78,7 +78,15 @@ const createSiteResourceSchema = z
         authDaemonMode: z.enum(["site", "remote", "native"]).optional(),
         pamMode: z.enum(["passthrough", "push"]).optional(),
         domainId: z.string().optional(), // only used for http mode, we need this to verify the alias is unique within the org
-        subdomain: z.string().optional() // only used for http mode, we need this to verify the alias is unique within the org
+        subdomain: z.string().optional(), // only used for http mode, we need this to verify the alias is unique within the org
+        aiProviderId: z
+            .number()
+            .int()
+            .positive()
+            .optional()
+            .describe(
+                "For inference-mode site resources: the AI provider this resource proxies chat completions to."
+            )
     })
     .strict()
     .refine(
@@ -171,6 +179,9 @@ const createSiteResourceSchema = z
     )
     .refine(
         (data) => {
+            if (data.mode == "inference") {
+                return true;
+            }
             return (
                 (data.siteIds !== undefined && data.siteIds.length > 0) ||
                 data.siteId !== undefined
@@ -319,7 +330,8 @@ export async function createSiteResource(
             authDaemonMode,
             pamMode,
             domainId,
-            subdomain
+            subdomain,
+            aiProviderId
         } = parsedBody.data;
 
         // Backward compatibility: merge deprecated siteId into siteIds array
@@ -487,7 +499,11 @@ export async function createSiteResource(
                 .where(
                     and(
                         eq(siteResources.orgId, orgId),
-                        eq(siteResources.alias, alias.trim())
+                        eq(siteResources.alias, alias.trim()),
+                        ne(
+                            siteResources.requiresExitNodeConnection,
+                            mode == "inference"
+                        ) // exclude looking at the ones on exit nodes if this is an inference resource
                     )
                 )
                 .limit(1);
@@ -524,6 +540,7 @@ export async function createSiteResource(
         let aliasAddress: string | null = null;
         let releaseAliasLock: (() => Promise<void>) | null = null;
         if (mode === "host" || mode === "http" || mode === "ssh") {
+            // no alias address but we do have an alias for inference
             const { value, release } =
                 await getNextAvailableAliasAddress(orgId);
             aliasAddress = value;
@@ -533,21 +550,24 @@ export async function createSiteResource(
         let newSiteResource: SiteResource | undefined;
         try {
             await db.transaction(async (trx) => {
-                const [network] = await trx
+                let network: typeof networks.$inferSelect | undefined;
-                    .insert(networks)
+                if (mode !== "inference") {
-                    .values({
+                    [network] = await trx
-                        scope: "resource",
+                        .insert(networks)
-                        orgId: orgId
+                        .values({
-                    })
+                            scope: "resource",
-                    .returning();
+                            orgId: orgId
+                        })
+                        .returning();
 
-                if (!network) {
+                    if (!network) {
-                    return next(
+                        return next(
-                        createHttpError(
+                            createHttpError(
-                            HttpCode.INTERNAL_SERVER_ERROR,
+                                HttpCode.INTERNAL_SERVER_ERROR,
-                            `Failed to create network`
+                                `Failed to create network`
-                        )
+                            )
-                    );
+                        );
+                    }
                 }
 
                 let tcpPortRangeStringAdjusted = tcpPortRangeString;
@@ -566,7 +586,7 @@ export async function createSiteResource(
                     name,
                     mode,
                     ssl,
-                    networkId: network.networkId,
+                    networkId: network ? network.networkId : null,
                     destination: destination, // the ssh can be null
                     scheme,
                     destinationPort,
@@ -582,7 +602,9 @@ export async function createSiteResource(
                         (mode == "http" || mode == "ssh" ? true : false), // default to true for http resources, otherwise false
                     domainId,
                     subdomain: finalSubdomain,
-                    fullDomain
+                    fullDomain,
+                    requiresExitNodeConnection: mode === "inference", // in the future we might want to have different modes that do this
+                    aiProviderId: aiProviderId ?? null
                 };
                 if (isLicensedSshPam) {
                     if (authDaemonPort !== undefined)
@@ -600,11 +622,13 @@ export async function createSiteResource(
 
                 //////////////////// update the associations ////////////////////
 
-                for (const siteId of siteIds) {
+                if (network) {
-                    await trx.insert(siteNetworks).values({
+                    for (const siteId of siteIds) {
-                        siteId: siteId,
+                        await trx.insert(siteNetworks).values({
-                        networkId: network.networkId
+                            siteId: siteId,
-                    });
+                            networkId: network.networkId
+                        });
+                    }
                 }
 
                 const [adminRole] = await trx

server/routers/siteResource/index.ts

@@ -17,3 +17,7 @@ export * from "./setSiteResourceClients";
 export * from "./addClientToSiteResource";
 export * from "./batchAddClientToSiteResources";
 export * from "./removeClientFromSiteResource";
+export * from "./listSiteResourceAiModels";
+export * from "./setSiteResourceAiModels";
+export * from "./addAiModelToSiteResource";
+export * from "./removeAiModelFromSiteResource";

server/routers/siteResource/listSiteResourceAiModels.ts

@@ -0,0 +1,115 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import {
+    db,
+    siteResources,
+    siteResourceAiModels,
+    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 listSiteResourceAiModelsParamsSchema = z.strictObject({
+    siteResourceId: z.coerce.number().int().positive()
+});
+
+async function query(siteResourceId: number) {
+    return await db
+        .select({
+            modelId: aiModels.modelId,
+            modelKey: aiModels.modelKey,
+            name: aiModels.name,
+            enabled: aiModels.enabled
+        })
+        .from(siteResourceAiModels)
+        .innerJoin(
+            aiModels,
+            eq(siteResourceAiModels.modelId, aiModels.modelId)
+        )
+        .where(eq(siteResourceAiModels.siteResourceId, siteResourceId));
+}
+
+export type ListSiteResourceAiModelsResponse = {
+    models: NonNullable<Awaited<ReturnType<typeof query>>>;
+};
+
+registry.registerPath({
+    method: "get",
+    path: "/site-resource/{siteResourceId}/ai-models",
+    description:
+        "List the AI models a site resource is restricted to. An empty list means the site resource is not restricted and every enabled model on its linked AI provider is allowed.",
+    tags: [OpenAPITags.PrivateResource],
+    request: {
+        params: listSiteResourceAiModelsParamsSchema
+    },
+    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 listSiteResourceAiModels(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedParams = listSiteResourceAiModelsParamsSchema.safeParse(
+            req.params
+        );
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { siteResourceId } = parsedParams.data;
+
+        const [siteResource] = await db
+            .select()
+            .from(siteResources)
+            .where(eq(siteResources.siteResourceId, siteResourceId))
+            .limit(1);
+
+        if (!siteResource) {
+            return next(
+                createHttpError(HttpCode.NOT_FOUND, "Site resource not found")
+            );
+        }
+
+        const models = await query(siteResourceId);
+
+        return response<ListSiteResourceAiModelsResponse>(res, {
+            data: { models },
+            success: true,
+            error: false,
+            message: "Site resource AI models retrieved successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}

server/routers/siteResource/removeAiModelFromSiteResource.ts

@@ -0,0 +1,140 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { db, siteResources, siteResourceAiModels } 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";
+
+const removeAiModelFromSiteResourceBodySchema = z.strictObject({
+    modelId: z.int().positive()
+});
+
+const removeAiModelFromSiteResourceParamsSchema = z.strictObject({
+    siteResourceId: z.coerce.number().int().positive()
+});
+
+registry.registerPath({
+    method: "post",
+    path: "/site-resource/{siteResourceId}/ai-models/remove",
+    description:
+        "Remove a single AI model from a site resource's model restriction allow-list.",
+    tags: [OpenAPITags.PrivateResource],
+    request: {
+        params: removeAiModelFromSiteResourceParamsSchema,
+        body: {
+            content: {
+                "application/json": {
+                    schema: removeAiModelFromSiteResourceBodySchema
+                }
+            }
+        }
+    },
+    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 removeAiModelFromSiteResource(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedBody = removeAiModelFromSiteResourceBodySchema.safeParse(
+            req.body
+        );
+        if (!parsedBody.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedBody.error).toString()
+                )
+            );
+        }
+
+        const { modelId } = parsedBody.data;
+
+        const parsedParams =
+            removeAiModelFromSiteResourceParamsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { siteResourceId } = parsedParams.data;
+
+        const [siteResource] = await db
+            .select()
+            .from(siteResources)
+            .where(eq(siteResources.siteResourceId, siteResourceId))
+            .limit(1);
+
+        if (!siteResource) {
+            return next(
+                createHttpError(HttpCode.NOT_FOUND, "Site resource not found")
+            );
+        }
+
+        const existingEntry = await db
+            .select()
+            .from(siteResourceAiModels)
+            .where(
+                and(
+                    eq(siteResourceAiModels.siteResourceId, siteResourceId),
+                    eq(siteResourceAiModels.modelId, modelId)
+                )
+            );
+
+        if (existingEntry.length === 0) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    "Model not found in site resource's restriction list"
+                )
+            );
+        }
+
+        await db
+            .delete(siteResourceAiModels)
+            .where(
+                and(
+                    eq(siteResourceAiModels.siteResourceId, siteResourceId),
+                    eq(siteResourceAiModels.modelId, modelId)
+                )
+            );
+
+        return response(res, {
+            data: {},
+            success: true,
+            error: false,
+            message: "Model removed from site resource successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}

server/routers/siteResource/setSiteResourceAiModels.ts

@@ -0,0 +1,167 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import {
+    db,
+    siteResources,
+    siteResourceAiModels,
+    aiModels
+} from "@server/db";
+import { eq, and, inArray } 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 setSiteResourceAiModelsBodySchema = z.strictObject({
+    modelIds: z.array(z.int().positive())
+});
+
+const setSiteResourceAiModelsParamsSchema = z.strictObject({
+    siteResourceId: z.coerce.number().int().positive()
+});
+
+registry.registerPath({
+    method: "post",
+    path: "/site-resource/{siteResourceId}/ai-models",
+    description:
+        "Set the AI models a site resource is restricted to. This replaces all existing restrictions. Pass an empty array to remove the restriction (allow every enabled model on the linked provider).",
+    tags: [OpenAPITags.PrivateResource],
+    request: {
+        params: setSiteResourceAiModelsParamsSchema,
+        body: {
+            content: {
+                "application/json": {
+                    schema: setSiteResourceAiModelsBodySchema
+                }
+            }
+        }
+    },
+    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 setSiteResourceAiModels(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedBody = setSiteResourceAiModelsBodySchema.safeParse(
+            req.body
+        );
+        if (!parsedBody.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedBody.error).toString()
+                )
+            );
+        }
+
+        const { modelIds } = parsedBody.data;
+
+        const parsedParams = setSiteResourceAiModelsParamsSchema.safeParse(
+            req.params
+        );
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { siteResourceId } = parsedParams.data;
+
+        const [siteResource] = await db
+            .select()
+            .from(siteResources)
+            .where(eq(siteResources.siteResourceId, siteResourceId))
+            .limit(1);
+
+        if (!siteResource) {
+            return next(
+                createHttpError(HttpCode.NOT_FOUND, "Site resource not found")
+            );
+        }
+
+        if (modelIds.length > 0) {
+            if (!siteResource.aiProviderId) {
+                return next(
+                    createHttpError(
+                        HttpCode.BAD_REQUEST,
+                        "Site resource has no AI provider linked"
+                    )
+                );
+            }
+
+            const validModels = await db
+                .select({ modelId: aiModels.modelId })
+                .from(aiModels)
+                .where(
+                    and(
+                        inArray(aiModels.modelId, modelIds),
+                        eq(aiModels.providerId, siteResource.aiProviderId)
+                    )
+                );
+
+            if (validModels.length !== new Set(modelIds).size) {
+                return next(
+                    createHttpError(
+                        HttpCode.BAD_REQUEST,
+                        "One or more model IDs do not exist or do not belong to this site resource's AI provider"
+                    )
+                );
+            }
+        }
+
+        await db.transaction(async (trx) => {
+            await trx
+                .delete(siteResourceAiModels)
+                .where(
+                    eq(siteResourceAiModels.siteResourceId, siteResourceId)
+                );
+
+            if (modelIds.length > 0) {
+                await trx
+                    .insert(siteResourceAiModels)
+                    .values(
+                        modelIds.map((modelId) => ({
+                            siteResourceId,
+                            modelId
+                        }))
+                    );
+            }
+        });
+
+        return response(res, {
+            data: {},
+            success: true,
+            error: false,
+            message: "AI models set for site resource successfully",
+            status: HttpCode.CREATED
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}

server/routers/siteResource/updateSiteResource.ts

@@ -50,7 +50,7 @@ const updateSiteResourceSchema = z
             )
             .optional(),
         // mode: z.enum(["host", "cidr", "port"]).optional(),
-        mode: z.enum(["host", "cidr", "http", "ssh"]).optional(),
+        mode: z.enum(["host", "cidr", "http", "ssh", "inference"]).optional(),
         ssl: z.boolean().optional(),
         scheme: z.enum(["http", "https"]).nullish(),
         destinationPort: z.int().positive().nullish(),
@@ -78,7 +78,16 @@ const updateSiteResourceSchema = z
         authDaemonMode: z.enum(["site", "remote", "native"]).optional(),
         pamMode: z.enum(["passthrough", "push"]).optional(),
         domainId: z.string().optional(),
-        subdomain: z.string().optional()
+        subdomain: z.string().optional(),
+        aiProviderId: z
+            .number()
+            .int()
+            .positive()
+            .nullable()
+            .optional()
+            .describe(
+                "For inference-mode site resources: the AI provider this resource proxies chat completions to. Set to null to unlink."
+            )
     })
     .strict()
     .refine(
@@ -173,6 +182,9 @@ const updateSiteResourceSchema = z
     )
     .refine(
         (data) => {
+            if (data.mode == "inference") {
+                return true;
+            }
             // if neither is provided, the existing site associations are left unchanged
             if (data.siteIds === undefined && data.siteId === undefined) {
                 return true;
@@ -326,7 +338,8 @@ export async function updateSiteResource(
             authDaemonMode,
             pamMode,
             domainId,
-            subdomain
+            subdomain,
+            aiProviderId
         } = parsedBody.data;
 
         // Backward compatibility: merge deprecated siteId into siteIds array
@@ -504,7 +517,11 @@ export async function updateSiteResource(
                     and(
                         eq(siteResources.orgId, existingSiteResource.orgId),
                         eq(siteResources.alias, alias.trim()),
-                        ne(siteResources.siteResourceId, siteResourceId) // exclude self
+                        ne(siteResources.siteResourceId, siteResourceId), // exclude self
+                        ne(
+                            siteResources.requiresExitNodeConnection,
+                            mode == "inference"
+                        ) // exclude looking at the ones on exit nodes if this is an inference resource
                     )
                 )
                 .limit(1);
@@ -579,13 +596,15 @@ export async function updateSiteResource(
                     disableIcmp:
                         mode !== undefined
                             ? disableIcmp ||
-                              (mode == "http" || mode == "ssh"
+                              (mode == "http" || mode == "ssh" ? true : false)
-                                  ? true
-                                  : false)
                             : disableIcmp,
                     domainId,
                     subdomain: finalSubdomain,
                     fullDomain,
+                    networkId: mode === "inference" ? null : undefined,
+                    requiresExitNodeConnection:
+                        mode !== undefined ? mode === "inference" : undefined,
+                    aiProviderId: aiProviderId,
                     ...sshPamSet
                 })
                 .where(and(eq(siteResources.siteResourceId, siteResourceId)))
@@ -593,7 +612,20 @@ export async function updateSiteResource(
 
             //////////////////// update the associations ////////////////////
 
-            if (siteIds !== undefined) {
+            if (mode === "inference") {
+                // inference resources are not attached to any site network
+                if (existingSiteResource.networkId) {
+                    await trx
+                        .delete(siteNetworks)
+                        .where(
+                            eq(
+                                siteNetworks.networkId,
+                                existingSiteResource.networkId
+                            )
+                        );
+                }
+                updatedSiteIds = [];
+            } else if (siteIds !== undefined) {
                 // delete the site - site resources associations
                 await trx
                     .delete(siteNetworks)

server/routers/target/createTarget.ts

@@ -6,7 +6,14 @@ import {
     TargetHealthCheck,
     targetHealthCheck
 } from "@server/db";
-import { newts, resources, sites, Target, targets } from "@server/db";
+import {
+    aiProviders,
+    newts,
+    resources,
+    sites,
+    Target,
+    targets
+} from "@server/db";
 import response from "@server/lib/response";
 import HttpCode from "@server/types/HttpCode";
 import createHttpError from "http-errors";
@@ -29,10 +36,19 @@ import { generateId } from "@server/auth/sessions/app";
 import config from "@server/lib/config";
 import { sendBrowserGatewayTargets } from "@server/routers/newt/targets";
 
-const createTargetParamsSchema = z.strictObject({
+const resourceTargetParamsSchema = z.strictObject({
     resourceId: z.coerce.number().int().positive()
 });
 
+const providerTargetParamsSchema = z.strictObject({
+    providerId: z.coerce.number().int().positive()
+});
+
+const createTargetParamsSchema = z.union([
+    resourceTargetParamsSchema,
+    providerTargetParamsSchema
+]);
+
 const createTargetSchema = z
     .strictObject({
         siteId: z.int().positive(),
@@ -95,7 +111,7 @@ registry.registerPath({
     description: "Create a target for a resource.",
     tags: [OpenAPITags.PublicResourceLegacy],
     request: {
-        params: createTargetParamsSchema,
+        params: resourceTargetParamsSchema,
         body: {
             content: {
                 "application/json": {
@@ -128,7 +144,40 @@ registry.registerPath({
     description: "Create a target for a resource.",
     tags: [OpenAPITags.PublicResource, OpenAPITags.Target],
     request: {
-        params: createTargetParamsSchema,
+        params: resourceTargetParamsSchema,
+        body: {
+            content: {
+                "application/json": {
+                    schema: createTargetSchema
+                }
+            }
+        }
+    },
+    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()
+                    })
+                }
+            }
+        }
+    }
+});
+
+registry.registerPath({
+    method: "put",
+    path: "/ai-provider/{providerId}/target",
+    description: "Create a target for an AI provider.",
+    tags: [OpenAPITags.AiProvider],
+    request: {
+        params: providerTargetParamsSchema,
         body: {
             content: {
                 "application/json": {
@@ -183,21 +232,74 @@ export async function createTarget(
             );
         }
 
-        const { resourceId } = parsedParams.data;
+        let resource: typeof resources.$inferSelect | undefined;
+        let provider: typeof aiProviders.$inferSelect | undefined;
 
-        // get the resource
+        if ("providerId" in parsedParams.data) {
-        const [resource] = await db
+            const { providerId } = parsedParams.data;
-            .select()
+            [provider] =
-            .from(resources)
+                req.aiProvider && req.aiProvider.providerId === providerId
-            .where(eq(resources.resourceId, resourceId));
+                    ? [req.aiProvider]
+                    : await db
+                          .select()
+                          .from(aiProviders)
+                          .where(eq(aiProviders.providerId, providerId))
+                          .limit(1);
 
-        if (!resource) {
+            if (!provider) {
-            return next(
+                return next(
-                createHttpError(
+                    createHttpError(
-                    HttpCode.NOT_FOUND,
+                        HttpCode.NOT_FOUND,
-                    `Resource with ID ${resourceId} not found`
+                        `AI provider with ID ${providerId} not found`
-                )
+                    )
-            );
+                );
+            }
+
+            if (provider.routingMode !== "target") {
+                return next(
+                    createHttpError(
+                        HttpCode.BAD_REQUEST,
+                        "AI provider must use target routing mode"
+                    )
+                );
+            }
+
+            if (provider.type !== "custom") {
+                return next(
+                    createHttpError(
+                        HttpCode.BAD_REQUEST,
+                        "Only custom AI providers support targets"
+                    )
+                );
+            }
+
+            if (
+                targetData.method &&
+                !["http", "https"].includes(targetData.method.toLowerCase())
+            ) {
+                return next(
+                    createHttpError(
+                        HttpCode.BAD_REQUEST,
+                        "AI provider target method must be http or https"
+                    )
+                );
+            }
+        } else {
+            const { resourceId } = parsedParams.data;
+            [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`
+                    )
+                );
+            }
         }
 
         const siteId = targetData.siteId;
@@ -217,6 +319,24 @@ export async function createTarget(
             );
         }
 
+        if (provider && site.orgId && site.orgId !== provider.orgId) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    "Site must belong to the AI provider organization"
+                )
+            );
+        }
+
+        const resourceId = resource?.resourceId ?? null;
+        const providerId = provider?.providerId ?? null;
+        const targetMode = provider
+            ? "http"
+            : (targetData.mode ?? resource?.mode ?? "http");
+        const targetMethod = provider
+            ? (targetData.method?.toLowerCase() ?? "https")
+            : targetData.method;
+
         const plainToken = generateId(48);
         const encryptedToken = encrypt(
             plainToken,
@@ -230,20 +350,24 @@ export async function createTarget(
             const existingTargets = await trx
                 .select()
                 .from(targets)
-                .where(eq(targets.resourceId, resourceId));
+                .where(
+                    providerId
+                        ? eq(targets.providerId, providerId)
+                        : eq(targets.resourceId, resourceId!)
+                );
 
             const existingTarget = existingTargets.find(
                 (target) =>
                     target.ip === targetData.ip &&
                     target.port === targetData.port &&
-                    target.method === targetData.method &&
+                    target.method === targetMethod &&
                     target.siteId === targetData.siteId
             );
 
             if (existingTarget) {
                 // log a warning
                 logger.warn(
-                    `Target with IP ${targetData.ip}, port ${targetData.port}, method ${targetData.method} already exists for resource ID ${resourceId}`
+                    `Target with IP ${targetData.ip}, port ${targetData.port}, method ${targetMethod} already exists for ${providerId ? `AI provider ID ${providerId}` : `resource ID ${resourceId}`}`
                 );
             }
 
@@ -252,10 +376,10 @@ export async function createTarget(
                     .insert(targets)
                     .values({
                         resourceId,
+                        providerId,
                         ...targetData,
-                        mode: (targetData.mode ??
+                        mode: targetMode as Target["mode"],
-                            resource.mode ??
+                        method: targetMethod,
-                            "http") as Target["mode"],
                         priority: targetData.priority || 100
                     })
                     .returning();
@@ -263,7 +387,7 @@ export async function createTarget(
                 // make sure the target is within the site subnet
                 if (
                     site.type == "wireguard" &&
-                    !isIpInCidr(targetData.ip, site.subnet!)
+                    !isIpInCidr(targetData.ip, site.exitNodeSubnet!)
                 ) {
                     return next(
                         createHttpError(
@@ -289,13 +413,12 @@ export async function createTarget(
                     .insert(targets)
                     .values({
                         resourceId,
+                        providerId,
                         siteId: site.siteId,
                         ip: targetData.ip,
-                        mode: (targetData.mode ??
+                        mode: targetMode as Target["mode"],
-                            resource.mode ??
-                            "http") as Target["mode"],
                         authToken: encryptedToken,
-                        method: targetData.method,
+                        method: targetMethod,
                         port: targetData.port,
                         internalPort,
                         enabled: targetData.enabled,
@@ -321,10 +444,12 @@ export async function createTarget(
             healthCheck = await trx
                 .insert(targetHealthCheck)
                 .values({
-                    orgId: resource.orgId,
+                    orgId: provider?.orgId ?? resource!.orgId,
                     targetId: newTarget[0].targetId,
                     siteId: targetData.siteId,
-                    name: `Resource ${resource.name} - ${targetData.ip}:${targetData.port}`,
+                    name: provider
+                        ? `AI Provider ${provider.name} - ${targetData.ip}:${targetData.port}`
+                        : `Resource ${resource!.name} - ${targetData.ip}:${targetData.port}`,
                     hcEnabled: targetData.hcEnabled ?? false,
                     hcPath: targetData.hcPath ?? null,
                     hcScheme: targetData.hcScheme ?? null,
@@ -399,10 +524,17 @@ export async function createTarget(
                         newt.newtId,
                         newTarget,
                         healthCheck,
-                        resource.mode === "udp" ? "udp" : "tcp",
+                        provider
+                            ? "tcp"
+                            : (resource!.mode as string) === "udp"
+                              ? "udp"
+                              : "tcp",
                         newt.version
                     );
-                } else if (["ssh", "rdp", "vnc"].includes(newTarget[0].mode)) {
+                } else if (
+                    !provider &&
+                    ["ssh", "rdp", "vnc"].includes(newTarget[0].mode)
+                ) {
                     await sendBrowserGatewayTargets(
                         newt.newtId,
                         newTarget,

server/routers/target/deleteTarget.ts

@@ -79,38 +79,54 @@ export async function deleteTarget(
                 )
             );
         }
-        // get the resource
-        const [resource] = await db
-            .select()
-            .from(resources)
-            .where(eq(resources.resourceId, deletedTarget.resourceId!));
 
-        if (!resource) {
+        if (
+            (!deletedTarget.resourceId && !deletedTarget.providerId) ||
+            (deletedTarget.resourceId && deletedTarget.providerId)
+        ) {
             return next(
                 createHttpError(
-                    HttpCode.NOT_FOUND,
+                    HttpCode.INTERNAL_SERVER_ERROR,
-                    `Resource with ID ${deletedTarget.resourceId} not found`
+                    `Target with ID ${targetId} has invalid ownership`
                 )
             );
         }
 
-        // check if there are other targets on the resource
+        let resource: typeof resources.$inferSelect | undefined;
-        const otherTargets = await db
+        if (deletedTarget.resourceId) {
-            .select()
+            [resource] = await db
-            .from(targets)
+                .select()
-            .where(
+                .from(resources)
-                and(
+                .where(eq(resources.resourceId, deletedTarget.resourceId))
-                    eq(targets.resourceId, resource.resourceId),
+                .limit(1);
-                    ne(targets.targetId, targetId)
-                )
-            );
 
-        if (otherTargets.length == 0) {
+            if (!resource) {
-            // set the resource status
+                return next(
-            await db
+                    createHttpError(
-                .update(resources)
+                        HttpCode.NOT_FOUND,
-                .set({ health: "unknown" })
+                        `Resource with ID ${deletedTarget.resourceId} not found`
-                .where(eq(resources.resourceId, resource.resourceId));
+                    )
+                );
+            }
+
+            // check if there are other targets on the resource
+            const otherTargets = await db
+                .select()
+                .from(targets)
+                .where(
+                    and(
+                        eq(targets.resourceId, resource.resourceId),
+                        ne(targets.targetId, targetId)
+                    )
+                );
+
+            if (otherTargets.length == 0) {
+                // set the resource status
+                await db
+                    .update(resources)
+                    .set({ health: "unknown" })
+                    .where(eq(resources.resourceId, resource.resourceId));
+            }
         }
 
         const [site] = await db
@@ -137,16 +153,26 @@ export async function deleteTarget(
                     .where(eq(newts.siteId, site.siteId))
                     .limit(1);
 
-                if (["http", "tcp", "udp"].includes(deletedTarget.mode)) {
+                if (
+                    deletedTarget.providerId ||
+                    ["http", "tcp", "udp"].includes(deletedTarget.mode)
+                ) {
                     await removeTargets(
                         newt.newtId,
                         // [deletedTarget],
                         [], // deleting the target from newt causes issues because we cant unbind the port. this needs to be fixed in newt before we can do this
                         [deletedHealthCheck],
-                        resource.mode === "udp" ? "udp" : "tcp",
+                        deletedTarget.providerId
+                            ? "tcp"
+                            : (resource!.mode as string) === "udp"
+                              ? "udp"
+                              : "tcp",
                         newt.version
                     );
-                } else if (["ssh", "rdp", "vnc"].includes(deletedTarget.mode)) {
+                } else if (
+                    !deletedTarget.providerId &&
+                    ["ssh", "rdp", "vnc"].includes(deletedTarget.mode)
+                ) {
                     await removeBrowserGatewayTarget(
                         newt.newtId,
                         deletedTarget.targetId,

server/routers/target/listTargets.ts

@@ -10,10 +10,19 @@ import { fromError } from "zod-validation-error";
 import logger from "@server/logger";
 import { OpenAPITags, registry } from "@server/openApi";
 
-const listTargetsParamsSchema = z.strictObject({
+const resourceTargetsParamsSchema = z.strictObject({
     resourceId: z.coerce.number().int().positive()
 });
 
+const providerTargetsParamsSchema = z.strictObject({
+    providerId: z.coerce.number().int().positive()
+});
+
+const listTargetsParamsSchema = z.union([
+    resourceTargetsParamsSchema,
+    providerTargetsParamsSchema
+]);
+
 const listTargetsSchema = z.strictObject({
     limit: z
         .string()
@@ -29,7 +38,7 @@ const listTargetsSchema = z.strictObject({
         .pipe(z.int().nonnegative())
 });
 
-function queryTargets(resourceId: number) {
+function queryTargets(owner: { resourceId: number } | { providerId: number }) {
     const baseQuery = db
         .select({
             targetId: targets.targetId,
@@ -39,6 +48,7 @@ function queryTargets(resourceId: number) {
             port: targets.port,
             enabled: targets.enabled,
             resourceId: targets.resourceId,
+            providerId: targets.providerId,
             siteId: targets.siteId,
             siteType: sites.type,
             siteName: sites.name,
@@ -71,7 +81,11 @@ function queryTargets(resourceId: number) {
             targetHealthCheck,
             eq(targetHealthCheck.targetId, targets.targetId)
         )
-        .where(eq(targets.resourceId, resourceId));
+        .where(
+            "providerId" in owner
+                ? eq(targets.providerId, owner.providerId)
+                : eq(targets.resourceId, owner.resourceId)
+        );
 
     return baseQuery;
 }
@@ -94,7 +108,7 @@ registry.registerPath({
     description: "List targets for a resource.",
     tags: [OpenAPITags.PublicResourceLegacy],
     request: {
-        params: listTargetsParamsSchema,
+        params: resourceTargetsParamsSchema,
         query: listTargetsSchema
     },
     responses: {
@@ -121,7 +135,34 @@ registry.registerPath({
     description: "List targets for a resource.",
     tags: [OpenAPITags.PublicResource, OpenAPITags.Target],
     request: {
-        params: listTargetsParamsSchema,
+        params: resourceTargetsParamsSchema,
+        query: listTargetsSchema
+    },
+    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()
+                    })
+                }
+            }
+        }
+    }
+});
+
+registry.registerPath({
+    method: "get",
+    path: "/ai-provider/{providerId}/targets",
+    description: "List targets for an AI provider.",
+    tags: [OpenAPITags.AiProvider],
+    request: {
+        params: providerTargetsParamsSchema,
         query: listTargetsSchema
     },
     responses: {
@@ -168,14 +209,18 @@ export async function listTargets(
                 )
             );
         }
-        const { resourceId } = parsedParams.data;
+        const owner = parsedParams.data;
+        const ownerCondition =
+            "providerId" in owner
+                ? eq(targets.providerId, owner.providerId)
+                : eq(targets.resourceId, owner.resourceId);
 
-        const baseQuery = queryTargets(resourceId);
+        const baseQuery = queryTargets(owner);
 
         const countQuery = db
             .select({ count: sql<number>`cast(count(*) as integer)` })
             .from(targets)
-            .where(eq(targets.resourceId, resourceId));
+            .where(ownerCondition);
 
         const targetsList = await baseQuery.limit(limit).offset(offset);
         const totalCountResult = await countQuery;

server/routers/target/updateTarget.ts

@@ -1,7 +1,7 @@
 import { Request, Response, NextFunction } from "express";
 import { z } from "zod";
 import { db, targetHealthCheck } from "@server/db";
-import { newts, resources, sites, targets } from "@server/db";
+import { aiProviders, newts, resources, sites, targets } from "@server/db";
 import { eq } from "drizzle-orm";
 import response from "@server/lib/response";
 import HttpCode from "@server/types/HttpCode";
@@ -147,21 +147,68 @@ export async function updateTarget(
             );
         }
 
-        // get the resource
+        if (
-        const [resource] = await db
+            (!target.resourceId && !target.providerId) ||
-            .select()
+            (target.resourceId && target.providerId)
-            .from(resources)
+        ) {
-            .where(eq(resources.resourceId, target.resourceId!));
-
-        if (!resource) {
             return next(
                 createHttpError(
-                    HttpCode.NOT_FOUND,
+                    HttpCode.INTERNAL_SERVER_ERROR,
-                    `Resource with ID ${target.resourceId} not found`
+                    `Target with ID ${targetId} has invalid ownership`
                 )
             );
         }
 
+        let resource: typeof resources.$inferSelect | undefined;
+        let provider: typeof aiProviders.$inferSelect | undefined;
+
+        if (target.resourceId) {
+            [resource] = await db
+                .select()
+                .from(resources)
+                .where(eq(resources.resourceId, target.resourceId))
+                .limit(1);
+
+            if (!resource) {
+                return next(
+                    createHttpError(
+                        HttpCode.NOT_FOUND,
+                        `Resource with ID ${target.resourceId} not found`
+                    )
+                );
+            }
+        } else {
+            [provider] = await db
+                .select()
+                .from(aiProviders)
+                .where(eq(aiProviders.providerId, target.providerId!))
+                .limit(1);
+
+            if (!provider) {
+                return next(
+                    createHttpError(
+                        HttpCode.NOT_FOUND,
+                        `AI provider with ID ${target.providerId} not found`
+                    )
+                );
+            }
+
+            if (
+                parsedBody.data.method !== undefined &&
+                (!parsedBody.data.method ||
+                    !["http", "https"].includes(
+                        parsedBody.data.method.toLowerCase()
+                    ))
+            ) {
+                return next(
+                    createHttpError(
+                        HttpCode.BAD_REQUEST,
+                        "AI provider target method must be http or https"
+                    )
+                );
+            }
+        }
+
         const [site] = await db
             .select()
             .from(sites)
@@ -177,6 +224,15 @@ export async function updateTarget(
             );
         }
 
+        if (provider && site.orgId && site.orgId !== provider.orgId) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    "Site must belong to the AI provider organization"
+                )
+            );
+        }
+
         const { internalPort, targetIps } = await pickPort(site.siteId!, db);
 
         if (!internalPort) {
@@ -221,8 +277,13 @@ export async function updateTarget(
         }
 
         const pathMatchTypeRemoved = parsedBody.data.pathMatchType === null;
-        const nextMode =
+        const nextMode = provider
-            parsedBody.data.mode === null ? undefined : parsedBody.data.mode;
+            ? parsedBody.data.mode !== undefined
+                ? "http"
+                : undefined
+            : parsedBody.data.mode === null
+              ? undefined
+              : parsedBody.data.mode;
 
         let updatedTarget: any;
         let updatedHc: any;
@@ -233,7 +294,10 @@ export async function updateTarget(
                     siteId: parsedBody.data.siteId,
                     ip: parsedBody.data.ip,
                     mode: nextMode,
-                    method: parsedBody.data.method,
+                    method:
+                        provider && parsedBody.data.method
+                            ? parsedBody.data.method.toLowerCase()
+                            : parsedBody.data.method,
                     port: parsedBody.data.port,
                     internalPort,
                     enabled: parsedBody.data.enabled,
@@ -368,15 +432,25 @@ export async function updateTarget(
                     .where(eq(newts.siteId, site.siteId))
                     .limit(1);
 
-                if (["http", "tcp", "udp"].includes(updatedTarget.mode)) {
+                if (
+                    provider ||
+                    ["http", "tcp", "udp"].includes(updatedTarget.mode)
+                ) {
                     await addTargets(
                         newt.newtId,
                         [updatedTarget],
                         [updatedHc],
-                        resource.mode === "udp" ? "udp" : "tcp",
+                        provider
+                            ? "tcp"
+                            : (resource!.mode as string) === "udp"
+                              ? "udp"
+                              : "tcp",
                         newt.version
                     );
-                } else if (["ssh", "rdp", "vnc"].includes(updatedTarget.mode)) {
+                } else if (
+                    !provider &&
+                    ["ssh", "rdp", "vnc"].includes(updatedTarget.mode)
+                ) {
                     await sendBrowserGatewayTargets(
                         newt.newtId,
                         [updatedTarget],

server/routers/traefik/traefikConfigProvider.ts

@@ -20,6 +20,9 @@ export async function traefikConfigProvider(
         const maintenancePort = config.getRawConfig().server.next_port;
         const maintenanceHost = config.getRawConfig().server.internal_hostname;
         const pangolinUIUrl = `http://${maintenanceHost}:${maintenancePort}`;
+        const aiGatewayUrl = `http://${maintenanceHost}:${
+            config.getRawConfig().server.internal_port
+        }/api/v1/ai-gateway`;
 
         const traefikConfig = await getTraefikConfig(
             currentExitNodeId,
@@ -28,7 +31,8 @@ export async function traefikConfigProvider(
             build != "oss", // generate the login pages on the cloud and and enterprise,
             config.getRawConfig().traefik.allow_raw_resources,
             pangolinUIUrl,
-            pangolinUIUrl
+            pangolinUIUrl,
+            aiGatewayUrl
         );
 
         if (traefikConfig?.http?.middlewares) {

server/routers/ws/messageHandlers.ts

@@ -5,7 +5,7 @@ import {
     handleNewtGetConfigMessage,
     handleDockerStatusMessage,
     handleDockerContainersMessage,
-    handleNewtPingRequestMessage,
+    handleNewtExitNodesRequestMessage,
     handleApplyBlueprintMessage,
     handleNewtPingMessage,
     startNewtOfflineChecker,
@@ -22,7 +22,8 @@ import {
     handleOlmDisconnectingMessage,
     handleOlmServerInitAddPeerHandshake,
     handleOlmLocalMessage,
-    handleOlmUnLocalMessage
+    handleOlmUnLocalMessage,
+    handleOlmExitNodesRequestMessage
 } from "../olm";
 import { handleHealthcheckStatusMessage } from "../target";
 import { handleRoundTripMessage } from "./handleRoundTripMessage";
@@ -37,6 +38,7 @@ export const messageHandlers: Record<string, MessageHandler> = {
     "olm/wg/local": handleOlmLocalMessage,
     "olm/wg/unlocal": handleOlmUnLocalMessage,
     "olm/ping": handleOlmPingMessage,
+    "olm/ping/request": handleOlmExitNodesRequestMessage,
     "olm/disconnecting": handleOlmDisconnectingMessage,
     "newt/disconnecting": handleNewtDisconnectingMessage,
     "newt/ping": handleNewtPingMessage,
@@ -45,7 +47,7 @@ export const messageHandlers: Record<string, MessageHandler> = {
     "newt/receive-bandwidth": handleReceiveBandwidthMessage,
     "newt/socket/status": handleDockerStatusMessage,
     "newt/socket/containers": handleDockerContainersMessage,
-    "newt/ping/request": handleNewtPingRequestMessage,
+    "newt/ping/request": handleNewtExitNodesRequestMessage,
     "newt/blueprint/apply": handleApplyBlueprintMessage,
     "newt/healthcheck/status": handleHealthcheckStatusMessage,
     "ws/round-trip/complete": handleRoundTripMessage

src/app/[orgId]/settings/clients/user/page.tsx

@@ -104,7 +104,9 @@ export default async function ClientsPage(props: ClientsPageProps) {
             archived: Boolean(client.archived),
             blocked: Boolean(client.blocked),
             approvalState: client.approvalState,
-            fingerprint
+            fingerprint,
+            firstSeen: client.firstSeen ?? null,
+            lastSeen: client.lastSeen ?? null
         };
     };
 

src/app/[orgId]/settings/resources/private/[niceId]/inference/page.tsx

@@ -0,0 +1,107 @@
+"use client";
+
+import {
+    SettingsContainer,
+    SettingsFormCell,
+    SettingsFormGrid,
+    SettingsSection,
+    SettingsSectionBody,
+    SettingsSectionDescription,
+    SettingsSectionFooter,
+    SettingsSectionForm,
+    SettingsSectionHeader,
+    SettingsSectionTitle
+} from "@app/components/Settings";
+import { Button } from "@app/components/ui/button";
+import { Form } from "@app/components/ui/form";
+import { createInferenceFormSchema } from "@app/lib/privateResourceForm";
+import { zodResolver } from "@hookform/resolvers/zod";
+import { useTranslations } from "next-intl";
+import { useActionState, useMemo, useState } from "react";
+import { useForm } from "react-hook-form";
+import { z } from "zod";
+import { PrivateResourceSitesField } from "@app/components/PrivateResourceSitesField";
+import { PrivateResourceInferenceDestinationFields } from "@app/components/PrivateResourceDestinationFields";
+import { PrivateResourcePortRanges } from "@app/components/PrivateResourcePortRanges";
+import { useSaveSiteResource } from "@app/hooks/useSaveSiteResource";
+import {
+    asAnyControl,
+    asAnySetValue,
+    asAnyWatch
+} from "@app/lib/formControlUtils";
+import { buildSelectedSitesForResource } from "@app/lib/privateResourceUtils";
+
+export default function PrivateResourceInferencePage() {
+    const t = useTranslations();
+    const { save, siteResource } = useSaveSiteResource();
+    const [selectedSites, setSelectedSites] = useState(() =>
+        buildSelectedSitesForResource(siteResource)
+    );
+
+    const formSchema = useMemo(() => createInferenceFormSchema(t), [t]);
+    type FormValues = z.infer<typeof formSchema>;
+
+    const form = useForm<FormValues>({
+        resolver: zodResolver(formSchema),
+        defaultValues: {
+            mode: "inference",
+            alias: siteResource.alias ?? null
+        }
+    });
+
+    const [, formAction, saveLoading] = useActionState(async () => {
+        const isValid = await form.trigger();
+        if (!isValid) return;
+
+        const data = form.getValues();
+        await save({
+            mode: "inference",
+            alias: data.alias
+        });
+    }, null);
+
+    return (
+        <SettingsContainer>
+            <SettingsSection>
+                <SettingsSectionHeader>
+                    <SettingsSectionTitle>
+                        {t("hostSettings")}
+                    </SettingsSectionTitle>
+                    <SettingsSectionDescription>
+                        {t("editInternalResourceDialogDestinationDescription")}
+                    </SettingsSectionDescription>
+                </SettingsSectionHeader>
+
+                <SettingsSectionBody>
+                    <SettingsSectionForm variant="half">
+                        <Form {...form}>
+                            <form
+                                action={formAction}
+                                id="private-resource-host-form"
+                            >
+                                <SettingsFormGrid>
+                                    <SettingsFormCell span="full">
+                                        <PrivateResourceInferenceDestinationFields
+                                            control={asAnyControl(form.control)}
+                                            watch={asAnyWatch(form.watch)}
+                                        />
+                                    </SettingsFormCell>
+                                </SettingsFormGrid>
+                            </form>
+                        </Form>
+                    </SettingsSectionForm>
+                </SettingsSectionBody>
+
+                <SettingsSectionFooter>
+                    <Button
+                        type="submit"
+                        form="private-resource-host-form"
+                        loading={saveLoading}
+                    >
+                        {t("saveSettings")}
+                    </Button>
+                </SettingsSectionFooter>
+            </SettingsSection>
+        </SettingsContainer>
+    );
+}

src/app/[orgId]/settings/resources/private/[niceId]/layout.tsx

@@ -52,7 +52,8 @@ export default async function PrivateResourceLayout(
         | "hostSettings"
         | "cidrSettings"
         | "httpSettings"
-        | "sshSettings";
+        | "sshSettings"
+        | "inferenceSettings";
 
     const navItems = [
         {

src/app/[orgId]/settings/resources/private/page.tsx

@@ -1,5 +1,5 @@
 import PrivateResourcesBanner from "@app/components/PrivateResourcesBanner";
-import type { InternalResourceRow } from "@app/components/PrivateResourcesTable";
+import type { PrivateResourceRow } from "@app/components/PrivateResourcesTable";
 import PrivateResourcesTable from "@app/components/PrivateResourcesTable";
 import SettingsSectionTitle from "@app/components/SettingsSectionTitle";
 import { internal } from "@app/lib/api";
@@ -61,7 +61,7 @@ export default async function ClientResourcesPage(
         redirect(`/${params.orgId}/settings/resources`);
     }
 
-    const internalResourceRows: InternalResourceRow[] = siteResources.map(
+    const internalResourceRows: PrivateResourceRow[] = siteResources.map(
         (siteResource) => {
             return {
                 id: siteResource.siteResourceId,

src/app/[orgId]/settings/resources/public/ProxyResourceTargetsForm.tsx

@@ -596,6 +596,7 @@ export function ProxyResourceTargetsForm({
             priority: 100,
             enabled: true,
             resourceId: resource?.resourceId ?? 0,
+            providerId: null,
             hcEnabled: false,
             hcPath: null,
             hcMethod: null,

src/app/[orgId]/settings/resources/public/create/page.tsx

@@ -206,7 +206,7 @@ function createAddTargetSchema(t: TranslateFn) {
         );
 }
 
-type NewResourceType = "http" | "ssh" | "rdp" | "vnc" | "tcp" | "udp";
+type NewResourceType = "http" | "ssh" | "rdp" | "vnc" | "tcp" | "udp" | "inference";
 
 type CreateBgTargetFormValues = SshSettingsFormValues;
 

src/app/[orgId]/settings/sites/[niceId]/credentials/page.tsx

@@ -72,8 +72,7 @@ export default function CredentialsPage() {
     const { data: latestVersions } = useQuery(
         productUpdatesQueries.latestVersion(true)
     );
-    const newtVersion =
+    const newtVersion = latestVersions?.data?.newt?.latestVersion ?? "latest";
-        latestVersions?.data?.newt?.latestVersion ?? "latest";
 
     // Fetch site defaults for wireguard sites to show in obfuscated config
     useEffect(() => {
@@ -354,7 +353,7 @@ export default function CredentialsPage() {
                                             text={generateObfuscatedWireGuardConfig(
                                                 {
                                                     subnet:
-                                                        site?.subnet ||
+                                                        site?.exitNodeSubnet ||
                                                         siteDefaults?.subnet ||
                                                         null,
                                                     address:

src/app/auth/login/page.tsx

@@ -91,6 +91,7 @@ export default async function Page(props: {
     let loginIdps: LoginFormIDP[] = [];
     let lastUsedIdpForSmartLogin: (LoginFormIDP & { orgId?: string }) | null =
         null;
+
     if (!useSmartLogin) {
         // Load IdPs for DashboardLoginForm (OSS or org-only IdP mode)
         if (build === "oss" || env.app.identityProviderMode !== "org") {
@@ -117,12 +118,12 @@ export default async function Page(props: {
                     `/idp/${persistedData.idpId}`
                 );
 
-                const idp = idpRes.data.data.idp;
+                const res = idpRes.data.data;
 
                 lastUsedIdpForSmartLogin = {
-                    idpId: idp.idpId,
+                    idpId: res.idp.idpId,
-                    name: idp.name,
+                    name: res.idp.name,
-                    variant: idp.type,
+                    variant: res.idpOidcConfig?.variant ?? res.idp.type,
                     orgId: persistedData.orgId,
                     lastUsed: true
                 };

src/components/IdpLoginButtons.tsx

@@ -114,59 +114,52 @@ export default function IdpLoginButtons({
 
             <div className="space-y-4">
                 {params.get("gotoapp") ? (
-                    <>
+                    <Button
-                        <Button
+                        type="button"
-                            type="button"
+                        className="w-full"
-                            className="w-full"
+                        onClick={() => {
-                            onClick={() => {
+                            goToApp();
-                                goToApp();
+                        }}
-                            }}
+                    >
-                        >
+                        {t("continueToApplication")}
-                            {t("continueToApplication")}
+                    </Button>
-                        </Button>
-                    </>
                 ) : (
-                    <>
+                    idps.map((idp) => {
-                        {idps.map((idp) => {
+                        const effectiveType =
-                            const effectiveType =
+                            idp.variant || idp.name.toLowerCase();
-                                idp.variant || idp.name.toLowerCase();
 
-                            return (
+                        return (
-                                <div
+                            <div className="w-full relative" key={idp.idpId}>
-                                    className="w-full relative"
+                                <Button
                                     key={idp.idpId}
+                                    type="button"
+                                    variant="outline"
+                                    className="w-full inline-flex items-center space-x-2  after:absolute after:inset-0 after:z-10"
+                                    onClick={() => {
+                                        startTransition(() =>
+                                            loginWithIdp(idp.idpId)
+                                        );
+                                    }}
+                                    disabled={loading}
+                                    loading={loading}
                                 >
-                                    <Button
+                                    <IdpTypeIcon
-                                        key={idp.idpId}
+                                        type={effectiveType}
-                                        type="button"
+                                        size={16}
-                                        variant="outline"
+                                    />
-                                        className="w-full inline-flex items-center space-x-2  after:absolute after:inset-0 after:z-10"
+                                    <span>{idp.name}</span>
-                                        onClick={() => {
+                                </Button>
-                                            startTransition(() =>
-                                                loginWithIdp(idp.idpId)
-                                            );
-                                        }}
-                                        disabled={loading}
-                                        loading={loading}
-                                    >
-                                        <IdpTypeIcon
-                                            type={effectiveType}
-                                            size={16}
-                                        />
-                                        <span>{idp.name}</span>
-                                    </Button>
 
-                                    {idp.lastUsed && (
+                                {idp.lastUsed && (
-                                        <div className="absolute inset-0">
+                                    <div className="absolute inset-0">
-                                            <span className="absolute top-0 right-0 text-xs bg-primary text-primary-foreground rounded-bl-sm rounded-tr-sm px-2 py-0.5">
+                                        <span className="absolute top-0 right-0 text-xs bg-primary text-primary-foreground rounded-bl-sm rounded-tr-sm px-2 py-0.5">
-                                                {t("idpLastUsed")}
+                                            {t("idpLastUsed")}
-                                            </span>
+                                        </span>
-                                        </div>
+                                    </div>
-                                    )}
+                                )}
-                                </div>
+                            </div>
-                            );
+                        );
-                        })}
+                    })
-                    </>
                 )}
             </div>
         </div>