feat(mcp): route hosted AI through OpenRouter with gpt-5.6-luna@medium

README.md

@@ -107,7 +107,8 @@ EMBEDDED_AGENT_PROVIDER=     # Required when multiple provider keys are set: 'op
 OPENAI_API_KEY=              # Required if using OpenAI
 ANTHROPIC_API_KEY=           # Required if using Anthropic
 OPENROUTER_API_KEY=          # Required if using OpenRouter
-OPENROUTER_MODEL=            # Optional OpenRouter model, defaults to 'openai/gpt-5'
+OPENROUTER_MODEL=            # Optional OpenRouter model, defaults to 'openai/gpt-5.6-luna'
+OPENROUTER_REASONING_EFFORT= # Optional OpenRouter reasoning effort, defaults to 'medium'
 
 # Optional overrides
 SENTRY_HOST=                 # For self-hosted deployments

docs/operations/embedded-agents.md

@@ -72,8 +72,9 @@ For OpenRouter:
 
 ```bash
 export OPENROUTER_API_KEY=sk-or-...
-# Optional; defaults to openai/gpt-5
-export OPENROUTER_MODEL=openai/gpt-5
+# Optional; defaults to openai/gpt-5.6-luna with reasoning effort medium
+export OPENROUTER_MODEL=openai/gpt-5.6-luna
+export OPENROUTER_REASONING_EFFORT=medium
 # Recommended, and required when multiple provider keys are set
 export EMBEDDED_AGENT_PROVIDER=openrouter
 ```
@@ -219,8 +220,9 @@ export OPENAI_MODEL=gpt-4
 # Anthropic (default: claude-opus-4-5-20251101)
 export ANTHROPIC_MODEL=claude-sonnet-4-5-20250929
 
-# OpenRouter (default: openai/gpt-5)
+# OpenRouter (default: openai/gpt-5.6-luna @ medium)
 export OPENROUTER_MODEL=anthropic/claude-sonnet-4
+export OPENROUTER_REASONING_EFFORT=high
 ```
 
 ### Verify Configuration

packages/mcp-cloudflare/package.json

@@ -46,7 +46,6 @@
   },
   "dependencies": {
     "@ai-sdk/mcp": "catalog:",
-    "@ai-sdk/openai": "catalog:",
     "@ai-sdk/react": "catalog:",
     "@cloudflare/workers-oauth-provider": "catalog:",
     "@modelcontextprotocol/sdk": "catalog:",

packages/mcp-cloudflare/src/server/lib/mcp-handler.test.ts

@@ -99,7 +99,9 @@ function createTestEnv(): Env {
     SENTRY_CLIENT_ID: "test-client-id",
     SENTRY_CLIENT_SECRET: "test-client-secret",
     SENTRY_HOST: "sentry.io",
-    OPENAI_API_KEY: "test-openai-key",
+    OPENROUTER_API_KEY: "test-openrouter-key",
+    OPENROUTER_MODEL: "openai/gpt-5.6-luna",
+    EMBEDDED_AGENT_PROVIDER: "openrouter",
     OAUTH_KV: {} as KVNamespace,
     OAUTH_PROVIDER: {
       listUserGrants: vi.fn().mockResolvedValue({ items: [] }),

packages/mcp-cloudflare/src/server/routes/chat.ts

@@ -1,6 +1,9 @@
 import { experimental_createMCPClient } from "@ai-sdk/mcp";
-import { openai } from "@ai-sdk/openai";
 import { StreamableHTTPClientTransport } from "@modelcontextprotocol/sdk/client/streamableHttp.js";
+import {
+  getOpenRouterModel,
+  getOpenRouterProviderOptions,
+} from "@sentry/mcp-core/internal/agents/openrouter-provider";
 import { logInfo, logIssue, logWarn } from "@sentry/mcp-core/telem/logging";
 import {
   convertToModelMessages,
@@ -93,9 +96,10 @@ async function refreshTokenIfNeeded(
 }
 
 export default new Hono<{ Bindings: Env }>().post("/", async (c) => {
-  // Validate that we have an OpenAI API key
-  if (!c.env.OPENAI_API_KEY) {
-    logIssue("OPENAI_API_KEY is not configured", {
+  // Validate that we have an OpenRouter API key (process.env is populated from
+  // worker bindings via nodejs_compat_populate_process_env).
+  if (!c.env.OPENROUTER_API_KEY && !process.env.OPENROUTER_API_KEY) {
+    logIssue("OPENROUTER_API_KEY is not configured", {
       loggerScope: ["cloudflare", "chat"],
     });
     return c.json(
@@ -307,7 +311,7 @@ export default new Hono<{ Bindings: Env }>().post("/", async (c) => {
     });
 
     const result = streamText({
-      model: openai("gpt-4o"),
+      model: getOpenRouterModel(),
       messages: modelMessages,
       tools,
       system: `You are an AI assistant designed EXCLUSIVELY for testing the Sentry MCP service. Your sole purpose is to help users test MCP functionality with their real Sentry account data - nothing more, nothing less.
@@ -332,8 +336,11 @@ Start conversations by exploring what's available in their account. Use tools li
 - \`get_sentry_resource\` to dive deep into a specific issue, event, or trace
 
 Remember: You're a test assistant, not a general-purpose helper. Stay focused on testing the MCP integration with their real data.`,
-      maxOutputTokens: 2000,
+      // Reasoning effort can consume completion budget before visible text/tool
+      // calls, so keep headroom above the old non-reasoning 2k cap.
+      maxOutputTokens: 16000,
       stopWhen: stepCountIs(10),
+      providerOptions: getOpenRouterProviderOptions(),
       experimental_telemetry: {
         isEnabled: true,
       },

packages/mcp-cloudflare/src/server/sentry.config.test.ts

@@ -13,7 +13,9 @@ function createEnv(overrides: Partial<Env> = {}): Env {
     SENTRY_CLIENT_SECRET: "test-client-secret",
     SENTRY_DSN: "https://public@example.ingest.sentry.io/1",
     SENTRY_HOST: "sentry.io",
-    OPENAI_API_KEY: "test-openai-key",
+    OPENROUTER_API_KEY: "test-openrouter-key",
+    OPENROUTER_MODEL: "openai/gpt-5.6-luna",
+    EMBEDDED_AGENT_PROVIDER: "openrouter",
     OAUTH_PROVIDER: {} as Env["OAUTH_PROVIDER"],
     AI: {} as Ai,
     ...overrides,

packages/mcp-cloudflare/src/server/types.ts

@@ -60,7 +60,12 @@ export interface Env {
   SENTRY_ENVIRONMENT?: string;
   SENTRY_DSN?: string;
   SENTRY_HOST?: string;
-  OPENAI_API_KEY: string;
+  /** @deprecated Prefer OPENROUTER_API_KEY for hosted MCP AI features. */
+  OPENAI_API_KEY?: string;
+  OPENROUTER_API_KEY?: string;
+  OPENROUTER_MODEL?: string;
+  OPENROUTER_REASONING_EFFORT?: string;
+  EMBEDDED_AGENT_PROVIDER?: string;
   MCP_URL?: string;
   OAUTH_PROVIDER: OAuthHelpers;
   AI: Ai;

packages/mcp-cloudflare/worker-configuration.d.ts

@@ -10,7 +10,11 @@ declare namespace Cloudflare {
 		SENTRY_CLIENT_ID: string;
 		SENTRY_CLIENT_SECRET: string;
 		SENTRY_DSN: string;
-		OPENAI_API_KEY: string;
+		OPENAI_API_KEY?: string;
+		OPENROUTER_API_KEY?: string;
+		OPENROUTER_MODEL?: string;
+		OPENROUTER_REASONING_EFFORT?: string;
+		EMBEDDED_AGENT_PROVIDER?: string;
 		COOKIE_SECRET: string;
 		CHAT_RATE_LIMITER: RateLimit;
 		SEARCH_RATE_LIMITER: RateLimit;
@@ -24,7 +28,7 @@ type StringifyValues<EnvType extends Record<string, unknown>> = {
 	[Binding in keyof EnvType]: EnvType[Binding] extends string ? EnvType[Binding] : string;
 };
 declare namespace NodeJS {
-	interface ProcessEnv extends StringifyValues<Pick<Cloudflare.Env, "SENTRY_CLIENT_ID" | "SENTRY_CLIENT_SECRET" | "SENTRY_DSN" | "OPENAI_API_KEY" | "COOKIE_SECRET">> {}
+	interface ProcessEnv extends StringifyValues<Pick<Cloudflare.Env, "SENTRY_CLIENT_ID" | "SENTRY_CLIENT_SECRET" | "SENTRY_DSN" | "OPENAI_API_KEY" | "OPENROUTER_API_KEY" | "OPENROUTER_MODEL" | "OPENROUTER_REASONING_EFFORT" | "EMBEDDED_AGENT_PROVIDER" | "COOKIE_SECRET">> {}
 }
 
 // Begin runtime types

packages/mcp-cloudflare/wrangler.test.jsonc

@@ -14,7 +14,9 @@
     "SENTRY_CLIENT_ID": "test-client-id",
     "SENTRY_CLIENT_SECRET": "test-client-secret",
     "SENTRY_HOST": "sentry.io",
-    "OPENAI_API_KEY": "test-openai-key"
+    "OPENROUTER_API_KEY": "test-openrouter-key",
+    "OPENROUTER_MODEL": "openai/gpt-5.6-luna",
+    "EMBEDDED_AGENT_PROVIDER": "openrouter"
   },
   "kv_namespaces": [
     {

packages/mcp-core/README.md

@@ -83,7 +83,8 @@ MCP_ADD_SCOPES=event:write             # Add to default scopes (keeps defaults)
 OPENAI_API_KEY=your-openai-key         # Use OpenAI for AI-powered search tools
 OPENAI_MODEL=gpt-5                     # OpenAI model to use (default: "gpt-5")
 OPENROUTER_API_KEY=your-openrouter-key # Or use OpenRouter for AI-powered search tools
-OPENROUTER_MODEL=openai/gpt-5          # OpenRouter model to use (default: "openai/gpt-5")
+OPENROUTER_MODEL=openai/gpt-5.6-luna   # OpenRouter model to use (default: "openai/gpt-5.6-luna")
+OPENROUTER_REASONING_EFFORT=medium     # OpenRouter reasoning effort: "none", "minimal", "low", "medium", "high", "xhigh" ("max" aliases xhigh), or "" to omit (default: "medium")
 OPENAI_REASONING_EFFORT=low            # Reasoning effort for o1 models: "low", "medium", "high", or "" to disable (default: "low")
 
 # No environment variable exists for the OpenAI base URL override; use --openai-base-url instead.

packages/mcp-core/src/internal/agents/openrouter-provider.test.ts

@@ -1,15 +1,21 @@
 import type { LanguageModelV3 } from "@ai-sdk/provider";
 import { generateText } from "ai";
 import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
-import { getOpenRouterModel } from "./openrouter-provider.js";
+import { ConfigurationError } from "../../errors.js";
+import {
+  getOpenRouterModel,
+  getOpenRouterProviderOptions,
+} from "./openrouter-provider.js";
 
 describe("openrouter-provider", () => {
   const originalApiKey = process.env.OPENROUTER_API_KEY;
   const originalModel = process.env.OPENROUTER_MODEL;
+  const originalReasoningEffort = process.env.OPENROUTER_REASONING_EFFORT;
 
   beforeEach(() => {
     process.env.OPENROUTER_API_KEY = "test-openrouter-key";
     delete process.env.OPENROUTER_MODEL;
+    delete process.env.OPENROUTER_REASONING_EFFORT;
   });
 
   afterEach(() => {
@@ -25,6 +31,12 @@ describe("openrouter-provider", () => {
       process.env.OPENROUTER_MODEL = originalModel;
     }
 
+    if (originalReasoningEffort === undefined) {
+      delete process.env.OPENROUTER_REASONING_EFFORT;
+    } else {
+      process.env.OPENROUTER_REASONING_EFFORT = originalReasoningEffort;
+    }
+
     vi.unstubAllGlobals();
   });
 
@@ -74,7 +86,7 @@ describe("openrouter-provider", () => {
 
   it("uses default and configured models", () => {
     expect((getOpenRouterModel() as LanguageModelV3).modelId).toBe(
-      "openai/gpt-5",
+      "openai/gpt-5.6-luna",
     );
 
     process.env.OPENROUTER_MODEL = "anthropic/claude-sonnet-4";
@@ -86,4 +98,57 @@ describe("openrouter-provider", () => {
       (getOpenRouterModel("google/gemini-2.5-pro") as LanguageModelV3).modelId,
     ).toBe("google/gemini-2.5-pro");
   });
+
+  it("defaults reasoning effort to medium and allows override", () => {
+    expect(getOpenRouterProviderOptions()).toEqual({
+      openai: {
+        structuredOutputs: false,
+        strictJsonSchema: false,
+        reasoningEffort: "medium",
+      },
+    });
+
+    process.env.OPENROUTER_REASONING_EFFORT = "high";
+
+    expect(getOpenRouterProviderOptions()).toEqual({
+      openai: {
+        structuredOutputs: false,
+        strictJsonSchema: false,
+        reasoningEffort: "high",
+      },
+    });
+
+    process.env.OPENROUTER_REASONING_EFFORT = "";
+
+    expect(getOpenRouterProviderOptions()).toEqual({
+      openai: {
+        structuredOutputs: false,
+        strictJsonSchema: false,
+      },
+    });
+  });
+
+  it("maps max to xhigh and rejects unknown reasoning effort", () => {
+    process.env.OPENROUTER_REASONING_EFFORT = "max";
+
+    expect(getOpenRouterProviderOptions()).toEqual({
+      openai: {
+        structuredOutputs: false,
+        strictJsonSchema: false,
+        reasoningEffort: "xhigh",
+      },
+    });
+
+    process.env.OPENROUTER_REASONING_EFFORT = "ludicrous";
+
+    expect(() => getOpenRouterProviderOptions()).toThrow(ConfigurationError);
+    expect(() => getOpenRouterProviderOptions()).toThrow(
+      /Invalid OPENROUTER_REASONING_EFFORT/,
+    );
+
+    // Prototype keys must not bypass validation via inherited Object props.
+    process.env.OPENROUTER_REASONING_EFFORT = "constructor";
+
+    expect(() => getOpenRouterProviderOptions()).toThrow(ConfigurationError);
+  });
 });

packages/mcp-core/src/internal/agents/openrouter-provider.ts

@@ -1,10 +1,64 @@
 import { createOpenAI } from "@ai-sdk/openai";
 import type { LanguageModel } from "ai";
+import { ConfigurationError } from "../../errors";
 import { USER_AGENT } from "../../version";
+import type { ProviderOptions } from "./types";
 
-const DEFAULT_OPENROUTER_MODEL = "openai/gpt-5";
+const DEFAULT_OPENROUTER_MODEL = "openai/gpt-5.6-luna";
+const DEFAULT_OPENROUTER_REASONING_EFFORT = "medium";
 const OPENROUTER_BASE_URL = "https://openrouter.ai/api/v1";
 
+// Matches @ai-sdk/openai chat providerOptions.openai.reasoningEffort.
+const OPENROUTER_REASONING_EFFORTS = [
+  "none",
+  "minimal",
+  "low",
+  "medium",
+  "high",
+  "xhigh",
+] as const;
+
+type OpenRouterReasoningEffort = (typeof OPENROUTER_REASONING_EFFORTS)[number];
+
+// Null prototype so Object.prototype keys like "constructor" cannot slip through.
+const OPENROUTER_REASONING_EFFORT_ALIASES: Record<
+  string,
+  OpenRouterReasoningEffort
+> = Object.assign(Object.create(null), {
+  // OpenRouter sometimes documents "max"; AI SDK accepts "xhigh".
+  max: "xhigh",
+});
+
+function resolveOpenRouterReasoningEffort(
+  value: string | undefined,
+): OpenRouterReasoningEffort | undefined {
+  if (value === undefined) {
+    return DEFAULT_OPENROUTER_REASONING_EFFORT;
+  }
+
+  // Empty string omits the provider option so the model default applies.
+  if (value === "") {
+    return undefined;
+  }
+
+  const normalized = value.trim().toLowerCase();
+  if (Object.hasOwn(OPENROUTER_REASONING_EFFORT_ALIASES, normalized)) {
+    return OPENROUTER_REASONING_EFFORT_ALIASES[normalized];
+  }
+
+  if (
+    (OPENROUTER_REASONING_EFFORTS as readonly string[]).includes(normalized)
+  ) {
+    return normalized as OpenRouterReasoningEffort;
+  }
+
+  throw new ConfigurationError(
+    `Invalid OPENROUTER_REASONING_EFFORT "${value}". Expected one of: ${OPENROUTER_REASONING_EFFORTS.join(
+      ", ",
+    )}, max (alias for xhigh), or "" to omit.`,
+  );
+}
+
 /**
  * Builds an OpenRouter chat-completions model for embedded agent calls.
  */
@@ -23,3 +77,29 @@ export function getOpenRouterModel(model?: string): LanguageModel {
     model ?? process.env.OPENROUTER_MODEL ?? DEFAULT_OPENROUTER_MODEL,
   );
 }
+
+/**
+ * Provider options for OpenRouter chat-completions calls.
+ *
+ * Uses the OpenAI chat provider options shape because OpenRouter is accessed
+ * through `@ai-sdk/openai` with an OpenRouter base URL.
+ */
+export function getOpenRouterProviderOptions(): ProviderOptions {
+  const reasoningEffort = resolveOpenRouterReasoningEffort(
+    process.env.OPENROUTER_REASONING_EFFORT,
+  );
+
+  return {
+    openai: {
+      // Required for optional structured-output fields.
+      // See: https://github.com/getsentry/sentry-mcp/issues/623
+      structuredOutputs: false,
+      strictJsonSchema: false,
+      ...(reasoningEffort
+        ? {
+            reasoningEffort,
+          }
+        : {}),
+    },
+  };
+}