first pass of traefik -> basic gateway

server/routers/aiGateway/chatCompletions.ts

@@ -0,0 +1,256 @@
+import { Request, Response } from "express";
+import { and, eq } from "drizzle-orm";
+import {
+    AiProvider,
+    aiModels,
+    aiProviders,
+    db,
+    resourceAiModels,
+    resources,
+    siteResourceAiModels,
+    siteResources
+} 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 logger from "@server/logger";
+import HttpCode from "@server/types/HttpCode";
+
+type ResolvedTarget = {
+    provider: AiProvider;
+    // null = no restriction; every enabled model on the provider is allowed
+    allowedModelIds: number[] | null;
+};
+
+async function resolveTarget(host: string): Promise<ResolvedTarget | null> {
+    const [resourceRow] = await db
+        .select({ resourceId: resources.resourceId, 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 {
+            provider: resourceRow.provider,
+            allowedModelIds: restrictions.length
+                ? restrictions.map((r) => r.modelId)
+                : null
+        };
+    }
+
+    const [siteResourceRow] = await db
+        .select({
+            siteResourceId: siteResources.siteResourceId,
+            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 {
+            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 } = target;
+        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/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/resource/createResource.ts

@@ -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/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/siteResource/createSiteResource.ts

@@ -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(
@@ -322,7 +330,8 @@ export async function createSiteResource(
             authDaemonMode,
             pamMode,
             domainId,
-            subdomain
+            subdomain,
+            aiProviderId
         } = parsedBody.data;
 
         // Backward compatibility: merge deprecated siteId into siteIds array
@@ -594,7 +603,8 @@ export async function createSiteResource(
                     domainId,
                     subdomain: finalSubdomain,
                     fullDomain,
-                    requiresExitNodeConnection: mode === "inference" // in the future we might want to have different modes that do this
+                    requiresExitNodeConnection: mode === "inference", // in the future we might want to have different modes that do this
+                    aiProviderId: aiProviderId ?? null
                 };
                 if (isLicensedSshPam) {
                     if (authDaemonPort !== undefined)

server/routers/siteResource/updateSiteResource.ts

@@ -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(
@@ -329,7 +338,8 @@ export async function updateSiteResource(
             authDaemonMode,
             pamMode,
             domainId,
-            subdomain
+            subdomain,
+            aiProviderId
         } = parsedBody.data;
 
         // Backward compatibility: merge deprecated siteId into siteIds array
@@ -594,6 +604,7 @@ export async function updateSiteResource(
                     networkId: mode === "inference" ? null : undefined,
                     requiresExitNodeConnection:
                         mode !== undefined ? mode === "inference" : undefined,
+                    aiProviderId: aiProviderId,
                     ...sshPamSet
                 })
                 .where(and(eq(siteResources.siteResourceId, siteResourceId)))

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) {