Cloudflare MCP Server Skill
Build and deploy Model Context Protocol (MCP) servers on Cloudflare Workers with TypeScript.
Status: Production Ready ✅ Last Updated: 2026-01-21 Latest Versions: @modelcontextprotocol/sdk@1.25.3, @cloudflare/workers-oauth-provider@0.2.2, agents@0.3.6
Recent Updates (2025):
-
September 2025: Code Mode (agents write code vs calling tools, auto-generated TypeScript API from schema)
-
August 2025: MCP Elicitation (interactive workflows, user input during execution), Task Queues, Email Integration
-
July 2025: MCPClientManager (connection management, OAuth flow, hibernation)
-
April 2025: HTTP Streamable Transport (single endpoint, recommended over SSE), Python MCP support
-
May 2025: Claude.ai remote MCP support, use-mcp React library, major partnerships
What is This Skill?
This skill teaches you to build remote MCP servers on Cloudflare - the ONLY platform with official remote MCP support.
Use when: Avoiding 24+ common MCP + Cloudflare errors (especially URL path mismatches - the #1 failure cause)
🚀 Quick Start (5 Minutes)
Start with Cloudflare's official template:
npm create cloudflare@latest -- my-mcp-server
--template=cloudflare/ai/demos/remote-mcp-authless
cd my-mcp-server && npm install && npm run dev
Choose template based on auth needs:
-
remote-mcp-authless
-
No auth (recommended for most)
-
remote-mcp-github-oauth
-
GitHub OAuth
-
remote-mcp-google-oauth
-
Google OAuth
-
remote-mcp-auth0 / remote-mcp-authkit
-
Enterprise SSO
-
mcp-server-bearer-auth
-
Custom auth
All templates: https://github.com/cloudflare/ai/tree/main/demos
Production examples: https://github.com/cloudflare/mcp-server-cloudflare (15 servers with real integrations)
Deployment Workflow
1. Create from template
npm create cloudflare@latest -- my-mcp --template=cloudflare/ai/demos/remote-mcp-authless cd my-mcp && npm install && npm run dev
2. Deploy
npx wrangler deploy
Note the output URL: https://my-mcp.YOUR_ACCOUNT.workers.dev
3. Test (PREVENTS 80% OF ERRORS!)
curl https://my-mcp.YOUR_ACCOUNT.workers.dev/sse
Expected: {"name":"My MCP Server","version":"1.0.0","transports":["/sse","/mcp"]}
Got 404? See "HTTP Transport Fundamentals" below
4. Configure client (~/.config/claude/claude_desktop_config.json)
{ "mcpServers": { "my-mcp": { "url": "https://my-mcp.YOUR_ACCOUNT.workers.dev/sse" // Must match curl URL! } } }
5. Restart Claude Desktop (config only loads at startup)
Post-Deployment Checklist:
-
curl returns server info (not 404)
-
Client URL matches curl URL exactly
-
Claude Desktop restarted
-
Tools visible in Claude Desktop
-
Test tool call succeeds
⚠️ CRITICAL: HTTP Transport Fundamentals
The #1 reason MCP servers fail to connect is URL path configuration mistakes.
URL Path Configuration Deep-Dive
When you serve an MCP server at a specific path, the client URL must match exactly.
Example 1: Serving at /sse
// src/index.ts export default { fetch(request: Request, env: Env, ctx: ExecutionContext) { const { pathname } = new URL(request.url);
if (pathname.startsWith("/sse")) {
return MyMCP.serveSSE("/sse").fetch(request, env, ctx); // ← Base path is "/sse"
}
return new Response("Not Found", { status: 404 });
} };
Client configuration MUST include /sse :
{ "mcpServers": { "my-mcp": { "url": "https://my-mcp.workers.dev/sse" // ✅ Correct } } }
❌ WRONG client configurations:
"url": "https://my-mcp.workers.dev" // Missing /sse → 404 "url": "https://my-mcp.workers.dev/" // Missing /sse → 404 "url": "http://localhost:8788" // Wrong after deploy
Example 2: Serving at / (root)
export default { fetch(request: Request, env: Env, ctx: ExecutionContext) { return MyMCP.serveSSE("/").fetch(request, env, ctx); // ← Base path is "/" } };
Client configuration:
{ "mcpServers": { "my-mcp": { "url": "https://my-mcp.workers.dev" // ✅ Correct (no /sse) } } }
How Base Path Affects Tool URLs
When you call serveSSE("/sse") , MCP tools are served at:
https://my-mcp.workers.dev/sse/tools/list https://my-mcp.workers.dev/sse/tools/call https://my-mcp.workers.dev/sse/resources/list
When you call serveSSE("/") , MCP tools are served at:
https://my-mcp.workers.dev/tools/list https://my-mcp.workers.dev/tools/call https://my-mcp.workers.dev/resources/list
The base path is prepended to all MCP endpoints automatically.
Request/Response Lifecycle
- Client connects to: https://my-mcp.workers.dev/sse ↓
- Worker receives request: { url: "https://my-mcp.workers.dev/sse", ... } ↓
- Your fetch handler: const { pathname } = new URL(request.url) ↓
- pathname === "/sse" → Check passes ↓
- MyMCP.serveSSE("/sse").fetch() → MCP server handles request ↓
- Tool calls routed to: /sse/tools/call
If client connects to https://my-mcp.workers.dev (missing /sse ):
pathname === "/" → Check fails → 404 Not Found
Testing Your URL Configuration
Step 1: Deploy your MCP server
npx wrangler deploy
Output: Deployed to https://my-mcp.YOUR_ACCOUNT.workers.dev
Step 2: Test the base path with curl
If serving at /sse, test this URL:
curl https://my-mcp.YOUR_ACCOUNT.workers.dev/sse
Should return MCP server info (not 404)
Step 3: Update client config with the EXACT URL you tested
{ "mcpServers": { "my-mcp": { "url": "https://my-mcp.YOUR_ACCOUNT.workers.dev/sse" // Match curl URL } } }
Step 4: Restart Claude Desktop
Post-Deployment Checklist
After deploying, verify:
-
curl https://worker.dev/sse returns MCP server info (not 404)
-
Client config URL matches deployed URL exactly
-
No typos in URL (common: workes.dev instead of workers.dev )
-
Using https:// (not http:// ) for deployed Workers
-
If using OAuth, redirect URI also updated
Transport Selection
Two transports available:
SSE (Server-Sent Events) - Legacy, wide compatibility
MyMCP.serveSSE("/sse").fetch(request, env, ctx)
Streamable HTTP - 2025 standard (recommended), single endpoint
MyMCP.serve("/mcp").fetch(request, env, ctx)
Support both for maximum compatibility:
export default { fetch(request: Request, env: Env, ctx: ExecutionContext) { const { pathname } = new URL(request.url);
if (pathname.startsWith("/sse")) {
return MyMCP.serveSSE("/sse").fetch(request, env, ctx);
}
if (pathname.startsWith("/mcp")) {
return MyMCP.serve("/mcp").fetch(request, env, ctx);
}
return new Response("Not Found", { status: 404 });
} };
CRITICAL: Use pathname.startsWith() to match paths correctly!
2025 Knowledge Gaps
MCP Elicitation (August 2025)
MCP servers can now request user input during tool execution:
// Request user input during tool execution const result = await this.elicit({ prompt: "Enter your API key:", type: "password" });
// Interactive workflows with Durable Objects state await this.state.storage.put("api_key", result);
Use cases: Confirmations, forms, multi-step workflows State: Preserved during agent hibernation
Code Mode (September 2025)
Agents SDK converts MCP schema → TypeScript API:
// Old: Direct tool calls await server.callTool("get_user", { id: "123" });
// New: Type-safe generated API const user = await api.getUser("123");
Benefits: Auto-generated doc comments, type safety, code completion
MCPClientManager (July 2025)
New class for MCP client capabilities:
import { MCPClientManager } from "agents/mcp";
const manager = new MCPClientManager(env); await manager.connect("https://external-mcp.com/sse"); // Auto-discovers tools, resources, prompts // Handles reconnection, OAuth flow, hibernation
Task Queues & Email (August 2025)
// Task queues for background jobs await this.queue.send({ task: "process_data", data });
// Email integration async onEmail(message: Email) { // Process incoming email const response = await this.generateReply(message); await this.sendEmail(response); }
HTTP Streamable Transport Details (April 2025)
Single endpoint replaces separate connection/messaging endpoints:
// Old: Separate endpoints /connect // Initialize connection /message // Send/receive messages
// New: Single streamable endpoint /mcp // All communication via HTTP streaming
Benefit: Simplified architecture, better performance
Security Considerations
PKCE Bypass Vulnerability (CRITICAL)
CVE: GHSA-qgp8-v765-qxx9 Severity: Critical Fixed in: @cloudflare/workers-oauth-provider@0.0.5
Problem: Earlier versions of the OAuth provider library had a critical vulnerability that completely bypassed PKCE protection, potentially allowing attackers to intercept authorization codes.
Action Required:
Check current version
npm list @cloudflare/workers-oauth-provider
Update if < 0.0.5
npm install @cloudflare/workers-oauth-provider@latest
Minimum Safe Version: @cloudflare/workers-oauth-provider@0.0.5 or later
Token Storage Best Practices
Always use encrypted storage for OAuth tokens:
// ✅ GOOD: workers-oauth-provider handles encryption automatically export default new OAuthProvider({ kv: (env) => env.OAUTH_KV, // Tokens stored encrypted // ... });
// ❌ BAD: Storing tokens in plain text await env.KV.put("access_token", token); // Security risk!
User-scoped KV keys prevent data leakage between users:
// ✅ GOOD: Namespace by user ID
await env.KV.put(user:${userId}:todos, data);
// ❌ BAD: Global namespace
await env.KV.put(todos, data); // Data visible to all users!
Authentication Patterns
Choose auth based on use case:
No Auth - Internal tools, dev (Template: remote-mcp-authless )
Bearer Token - Custom auth (Template: mcp-server-bearer-auth )
// Validate Authorization: Bearer <token> const token = request.headers.get("Authorization")?.replace("Bearer ", ""); if (!await validateToken(token, env)) { return new Response("Unauthorized", { status: 401 }); }
OAuth Proxy - GitHub/Google (Template: remote-mcp-github-oauth )
import { OAuthProvider, GitHubHandler } from "@cloudflare/workers-oauth-provider";
export default new OAuthProvider({ authorizeEndpoint: "/authorize", tokenEndpoint: "/token", defaultHandler: new GitHubHandler({ clientId: (env) => env.GITHUB_CLIENT_ID, clientSecret: (env) => env.GITHUB_CLIENT_SECRET, scopes: ["repo", "user:email"] }), kv: (env) => env.OAUTH_KV, apiHandlers: { "/sse": MyMCP.serveSSE("/sse") } });
⚠️ CRITICAL: All OAuth URLs (url, authorizationUrl, tokenUrl) must use same domain
Remote OAuth with DCR - Full OAuth provider (Template: remote-mcp-authkit )
Security levels: No Auth (⚠️) < Bearer (✅) < OAuth Proxy (✅✅) < Remote OAuth (✅✅✅)
Stateful MCP Servers (Durable Objects)
McpAgent extends Durable Objects for per-session state:
// Storage API await this.state.storage.put("key", "value"); const value = await this.state.storage.get<string>("key");
// Required wrangler.jsonc { "durable_objects": { "bindings": [{ "name": "MY_MCP", "class_name": "MyMCP" }] }, "migrations": [{ "tag": "v1", "new_classes": ["MyMCP"] }] // Required on first deploy! }
Critical: Migrations required on first deployment
Cost: Durable Objects now included in free tier (2025)
Architecture: Internal vs External Transports
Important: McpAgent uses different transports for client-facing vs internal communication.
Source: GitHub Issue #172
Transport Architecture
Client --- (SSE or HTTP) --> Worker --- (WebSocket) --> Durable Object
Client → Worker (External):
-
SSE transport: /sse endpoint
-
HTTP Streamable: /mcp endpoint
-
Client chooses transport
Worker → Durable Object (Internal):
-
Always WebSocket
-
Required by PartyServer (McpAgent's internal dependency)
-
Automatic upgrade, invisible to client
What This Means
-
SSE clients are fully supported - External interface can be SSE
-
WebSocket is mandatory for DO - Internal Worker-DO communication always uses WebSocket
-
This is not a limitation - It's an implementation detail of McpAgent's architecture
Example
export default { fetch(request: Request, env: Env, ctx: ExecutionContext) { const { pathname } = new URL(request.url);
// Client uses SSE
if (pathname.startsWith("/sse")) {
// ✅ Client → Worker: SSE
// ✅ Worker → DO: WebSocket (automatic)
return MyMCP.serveSSE("/sse").fetch(request, env, ctx);
}
return new Response("Not Found", { status: 404 });
} };
Key Takeaway: You can serve SSE to clients without worrying about the internal WebSocket requirement.
Common Patterns
Tool Return Format (CRITICAL)
Source: Stytch Blog - Building MCP Server with OAuth
All MCP tools must return this exact format:
this.server.tool( "my_tool", { /* schema */ }, async (params) => { // ✅ CORRECT: Return object with content array return { content: [ { type: "text", text: "Your result here" } ] };
// ❌ WRONG: Raw string
return "Your result here";
// ❌ WRONG: Plain object
return { result: "Your result here" };
} );
Common mistake: Returning raw strings or plain objects instead of proper MCP content format. This causes client parsing errors.
Conditional Tool Registration
Source: Cloudflare Blog - Building AI Agents
Dynamically add tools based on authenticated user:
export class MyMCP extends McpAgent<Env> { async init() { this.server = new McpServer({ name: "My MCP" });
// Base tools for all users
this.server.tool("public_tool", { /* schema */ }, async (params) => {
// Available to everyone
});
// Conditional tools based on user
const userId = this.props?.userId;
if (await this.isAdmin(userId)) {
this.server.tool("admin_tool", { /* schema */ }, async (params) => {
// Only available to admins
});
}
// Premium features
if (await this.isPremiumUser(userId)) {
this.server.tool("premium_feature", { /* schema */ }, async (params) => {
// Only for premium users
});
}
}
private async isAdmin(userId?: string): Promise<boolean> {
if (!userId) return false;
const userRole = await this.state.storage.get<string>(user:${userId}:role);
return userRole === "admin";
}
}
Use cases:
-
Feature flags per user
-
Premium vs free tier tools
-
Role-based access control (RBAC)
-
A/B testing new tools
Caching with DO Storage
async getCached<T>(key: string, ttlMs: number, fetchFn: () => Promise<T>): Promise<T> { const cached = await this.state.storage.get<{ data: T, timestamp: number }>(key); if (cached && Date.now() - cached.timestamp < ttlMs) { return cached.data; } const data = await fetchFn(); await this.state.storage.put(key, { data, timestamp: Date.now() }); return data; }
Rate Limiting
async rateLimit(key: string, maxRequests: number, windowMs: number): Promise<boolean> {
const requests = await this.state.storage.get<number[]>(ratelimit:${key}) || [];
const recentRequests = requests.filter(ts => Date.now() - ts < windowMs);
if (recentRequests.length >= maxRequests) return false;
recentRequests.push(Date.now());
await this.state.storage.put(ratelimit:${key}, recentRequests);
return true;
}
24 Known Errors (With Solutions)
- McpAgent Class Not Exported
Error: TypeError: Cannot read properties of undefined (reading 'serve')
Cause: Forgot to export McpAgent class
Solution:
export class MyMCP extends McpAgent { ... } // ✅ Must export export default { fetch() { ... } }
- Base Path Configuration Mismatch (Most Common!)
Error: 404 Not Found or Connection failed
Cause: serveSSE("/sse") but client configured with https://worker.dev (missing /sse )
Solution: Match base paths exactly
// Server serves at /sse MyMCP.serveSSE("/sse").fetch(...)
// Client MUST include /sse { "url": "https://worker.dev/sse" } // ✅ Correct { "url": "https://worker.dev" } // ❌ Wrong - 404
Debug steps:
-
Check what path your server uses: serveSSE("/sse") vs serveSSE("/")
-
Test with curl: curl https://worker.dev/sse
-
Update client config to match curl URL
- Transport Type Confusion
Error: Connection failed: Unexpected response format
Cause: Client expects SSE but connects to HTTP endpoint (or vice versa)
Solution: Match transport types
// SSE transport MyMCP.serveSSE("/sse") // Client URL: https://worker.dev/sse
// HTTP transport MyMCP.serve("/mcp") // Client URL: https://worker.dev/mcp
Best practice: Support both transports (see Transport Selection Guide)
- pathname.startsWith() Logic Error
Error: Both /sse and /mcp routes fail or conflict
Cause: Incorrect path matching logic
Solution: Use startsWith() correctly
// ✅ CORRECT if (pathname.startsWith("/sse")) { return MyMCP.serveSSE("/sse").fetch(...); } if (pathname.startsWith("/mcp")) { return MyMCP.serve("/mcp").fetch(...); }
// ❌ WRONG: Exact match breaks sub-paths if (pathname === "/sse") { // Breaks /sse/tools/list return MyMCP.serveSSE("/sse").fetch(...); }
- Local vs Deployed URL Mismatch
Error: Works in dev, fails after deployment
Cause: Client still configured with localhost URL
Solution: Update client config after deployment
// Development { "url": "http://localhost:8788/sse" }
// ⚠️ MUST UPDATE after npx wrangler deploy { "url": "https://my-mcp.YOUR_ACCOUNT.workers.dev/sse" }
Post-deployment checklist:
-
Run npx wrangler deploy and note output URL
-
Update client config with deployed URL
-
Test with curl
-
Restart Claude Desktop
- OAuth Redirect URI Mismatch
Error: OAuth error: redirect_uri does not match
Cause: OAuth redirect URI doesn't match deployed URL
Solution: Update ALL OAuth URLs after deployment
{ "url": "https://my-mcp.YOUR_ACCOUNT.workers.dev/sse", "auth": { "type": "oauth", "authorizationUrl": "https://my-mcp.YOUR_ACCOUNT.workers.dev/authorize", // Must match deployed domain "tokenUrl": "https://my-mcp.YOUR_ACCOUNT.workers.dev/token" } }
CRITICAL: All URLs must use the same protocol and domain!
- Missing OPTIONS Handler (CORS Preflight)
Error: Access to fetch at '...' blocked by CORS policy or Method Not Allowed
Cause: Browser clients send OPTIONS requests for CORS preflight, but server doesn't handle them
Solution: Add OPTIONS handler
export default { fetch(request: Request, env: Env, ctx: ExecutionContext) { // Handle CORS preflight if (request.method === "OPTIONS") { return new Response(null, { status: 204, headers: { "Access-Control-Allow-Origin": "*", "Access-Control-Allow-Methods": "GET, POST, OPTIONS", "Access-Control-Allow-Headers": "Content-Type, Authorization", "Access-Control-Max-Age": "86400" } }); }
// ... rest of your fetch handler
} };
When needed: Browser-based MCP clients (like MCP Inspector in browser)
- Request Body Validation Missing
Error: TypeError: Cannot read properties of undefined or Unexpected token in JSON parsing
Cause: Client sends malformed JSON, server doesn't validate before parsing
Solution: Wrap request handling in try/catch
export default { async fetch(request: Request, env: Env, ctx: ExecutionContext) { try { // Your MCP server logic return await MyMCP.serveSSE("/sse").fetch(request, env, ctx); } catch (error) { console.error("Request handling error:", error); return new Response( JSON.stringify({ error: "Invalid request", details: error.message }), { status: 400, headers: { "Content-Type": "application/json" } } ); } } };
- Environment Variable Validation Missing
Error: TypeError: env.API_KEY is undefined or silent failures (tools return empty data)
Cause: Required environment variables not configured or missing at runtime
Solution: Add startup validation
export class MyMCP extends McpAgent<Env> { async init() { // Validate required environment variables if (!this.env.API_KEY) { throw new Error("API_KEY environment variable not configured"); } if (!this.env.DATABASE_URL) { throw new Error("DATABASE_URL environment variable not configured"); }
// Continue with tool registration
this.server.tool(...);
} }
Configuration checklist:
-
Development: Add to .dev.vars (local only, gitignored)
-
Production: Add to wrangler.jsonc vars (public) or use wrangler secret (sensitive)
Best practices:
.dev.vars (local development, gitignored)
API_KEY=dev-key-123 DATABASE_URL=http://localhost:3000
wrangler.jsonc (public config)
{ "vars": { "ENVIRONMENT": "production", "LOG_LEVEL": "info" } }
wrangler secret (production secrets)
npx wrangler secret put API_KEY npx wrangler secret put DATABASE_URL
- McpAgent vs McpServer Confusion
Error: TypeError: server.registerTool is not a function or this.server is undefined
Cause: Trying to use standalone SDK patterns with McpAgent class
Solution: Use McpAgent's this.server.tool() pattern
// ❌ WRONG: Mixing standalone SDK with McpAgent import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
const server = new McpServer({ name: "My Server" }); server.registerTool(...); // Not compatible with McpAgent!
export class MyMCP extends McpAgent { /* no server property */ }
// ✅ CORRECT: McpAgent pattern export class MyMCP extends McpAgent<Env> { server = new McpServer({ name: "My MCP Server", version: "1.0.0" });
async init() { this.server.tool("tool_name", ...); // Use this.server } }
Key difference: McpAgent provides this.server property, standalone SDK doesn't.
- WebSocket Hibernation State Loss
Error: Tool calls fail after reconnect with "state not found"
Cause: In-memory state cleared on hibernation
Solution: Use this.state.storage instead of instance properties
// ❌ DON'T: Lost on hibernation this.userId = "123";
// ✅ DO: Persists through hibernation await this.state.storage.put("userId", "123");
- Durable Objects Binding Missing
Error: TypeError: Cannot read properties of undefined (reading 'idFromName')
Cause: Forgot DO binding in wrangler.jsonc
Solution: Add binding (see Stateful MCP Servers section)
{ "durable_objects": { "bindings": [ { "name": "MY_MCP", "class_name": "MyMCP", "script_name": "my-mcp-server" } ] } }
- Migration Not Defined
Error: Error: Durable Object class MyMCP has no migration defined
Cause: First DO deployment requires migration
Solution:
{ "migrations": [ { "tag": "v1", "new_classes": ["MyMCP"] } ] }
- serializeAttachment() Not Used
Error: WebSocket metadata lost on hibernation wake
Cause: Not using serializeAttachment() to preserve connection metadata
Solution: See WebSocket Hibernation section
- OAuth Consent Screen Disabled
Security risk: Users don't see what permissions they're granting
Cause: allowConsentScreen: false in production
Solution: Always enable in production
export default new OAuthProvider({ allowConsentScreen: true, // ✅ Always true in production // ... });
- JWT Signing Key Missing
Error: Error: JWT_SIGNING_KEY environment variable not set
Cause: OAuth Provider requires signing key for tokens
Solution:
Generate secure key
openssl rand -base64 32
Add to wrangler secret
npx wrangler secret put JWT_SIGNING_KEY
- Tool Schema Validation Error
Error: ZodError: Invalid input type
Cause: Client sends string, schema expects number (or vice versa)
Solution: Use Zod transforms
// Accept string, convert to number param: z.string().transform(val => parseInt(val, 10))
// Or: Accept both types param: z.union([z.string(), z.number()]).transform(val => typeof val === "string" ? parseInt(val, 10) : val )
- Multiple Transport Endpoints Conflicting
Error: /sse returns 404 after adding /mcp
Cause: Incorrect path matching (missing startsWith() )
Solution: Use startsWith() or exact matches correctly (see Error #4)
- Local Testing with Miniflare Limitations
Error: OAuth flow fails in local dev, or Durable Objects behave differently
Cause: Miniflare doesn't support all DO features
Solution: Use npx wrangler dev --remote for full DO support
Local simulation (faster but limited)
npm run dev
Remote DOs (slower but accurate)
npx wrangler dev --remote
- Client Configuration Format Error
Error: Claude Desktop doesn't recognize server
Cause: Wrong JSON format in claude_desktop_config.json
Solution: See "Connect Claude Desktop" section for correct format
Common mistakes:
// ❌ WRONG: Missing "mcpServers" wrapper { "my-mcp": { "url": "https://worker.dev/sse" } }
// ❌ WRONG: Trailing comma { "mcpServers": { "my-mcp": { "url": "https://worker.dev/sse", // ← Remove comma } } }
// ✅ CORRECT { "mcpServers": { "my-mcp": { "url": "https://worker.dev/sse" } } }
- Health Check Endpoint Missing
Issue: Can't tell if Worker is running or if URL is correct
Impact: Debugging connection issues takes longer
Solution: Add health check endpoint (see Transport Selection Guide)
Test:
curl https://my-mcp.workers.dev/health
Should return: {"status":"ok","transports":{...}}
- CORS Headers Missing
Error: Access to fetch at '...' blocked by CORS policy
Cause: MCP server doesn't return CORS headers for cross-origin requests
Solution: Add CORS headers to all responses
// Manual CORS (if not using OAuthProvider) const corsHeaders = { "Access-Control-Allow-Origin": "*", // Or specific origin "Access-Control-Allow-Methods": "GET, POST, OPTIONS", "Access-Control-Allow-Headers": "Content-Type, Authorization" };
// Add to responses return new Response(body, { headers: { ...corsHeaders, "Content-Type": "application/json" } });
Note: OAuthProvider handles CORS automatically!
- IoContext Timeout During MCP Initialization
Error: IoContext timed out due to inactivity, waitUntil tasks were cancelled
Source: GitHub Issue #640
Cause: When implementing MCP servers using McpAgent with custom Bearer authentication, the IoContext times out during the MCP protocol initialization handshake (before any tools are called).
Symptoms:
-
Timeout occurs before any tools are called
-
~2 minute gap between initial request and agent initialization
-
Internal methods work (setInitializeRequest, getInitializeRequest, updateProps)
-
Both GET and POST to /mcp are canceled
-
Error: "IoContext timed out due to inactivity, waitUntil tasks were cancelled"
Affected Code Pattern:
// Custom Bearer auth without OAuthProvider wrapper export default { fetch: async (req, env, ctx) => { const authHeader = req.headers.get("Authorization"); if (!authHeader?.startsWith("Bearer ")) { return new Response("Unauthorized", { status: 401 }); }
if (url.pathname === "/sse") {
return MyMCP.serveSSE("/sse")(req, env, ctx); // ← Timeout here
}
return new Response("Not found", { status: 404 });
} };
Root Cause Hypothesis:
-
May require OAuthProvider wrapper even for custom Bearer auth
-
Possible missing timeout configuration for Durable Object IoContext
-
May need CloudflareMCPServer instead of standard McpServer
Workaround: Use official templates with OAuthProvider pattern instead of custom Bearer auth:
// Use OAuthProvider wrapper (recommended) import { OAuthProvider } from "@cloudflare/workers-oauth-provider";
export default new OAuthProvider({ authorizeEndpoint: "/authorize", tokenEndpoint: "/token", // ... OAuth config apiHandlers: { "/sse": MyMCP.serveSSE("/sse") } });
Status: Investigation ongoing (issue open as of 2026-01-21)
- OAuth Remote Connection Failures
Error: Connection to remote MCP server fails when using OAuth (works locally but fails when deployed)
Source: GitHub Issue #444
Cause: When deploying MCP client from Cloudflare Agents repository to Workers, client fails to connect to MCP servers secured with OAuth.
Symptoms:
-
Works perfectly in local development
-
Fails after deployment to Workers
-
OAuth handshake never completes
-
Client can't establish connection
Troubleshooting Steps:
Verify OAuth tokens are handled correctly during remote connection attempts
// Check token is being passed to remote server console.log("Connecting with token:", token ? "present" : "missing");
Check network permissions to access OAuth provider
// Ensure Worker can reach OAuth endpoints const response = await fetch("https://oauth-provider.com/token");
Verify CORS configuration on OAuth provider
// OAuth provider must allow Worker origin headers: { "Access-Control-Allow-Origin": "https://your-worker.workers.dev", "Access-Control-Allow-Methods": "POST, OPTIONS", "Access-Control-Allow-Headers": "Content-Type, Authorization" }
Check redirect URIs match deployed URLs
{ "url": "https://mcp.workers.dev/sse", "auth": { "authorizationUrl": "https://mcp.workers.dev/authorize", // Must match deployed domain "tokenUrl": "https://mcp.workers.dev/token" } }
Deployment Checklist:
-
All OAuth URLs use deployed domain (not localhost)
-
CORS headers configured on OAuth provider
-
Network requests to OAuth provider allowed in Worker
-
Redirect URIs registered with OAuth provider
-
Environment variables set in production (wrangler secret )
Related: Issue #640 (both involve OAuth/auth in remote deployments)
Testing & Deployment
Local dev
npm run dev # Miniflare (fast) npx wrangler dev --remote # Remote DOs (accurate)
Test with MCP Inspector
npx @modelcontextprotocol/inspector@latest
Open http://localhost:5173, enter http://localhost:8788/sse
Deploy
npx wrangler login # First time only npx wrangler deploy
⚠️ CRITICAL: Update client config with deployed URL!
Monitor logs
npx wrangler tail
Official Documentation
-
Cloudflare Agents: https://developers.cloudflare.com/agents/
-
MCP Specification: https://modelcontextprotocol.io/
-
Official Templates: https://github.com/cloudflare/ai/tree/main/demos
-
Production Servers: https://github.com/cloudflare/mcp-server-cloudflare
-
workers-oauth-provider: https://github.com/cloudflare/workers-oauth-provider
Package Versions: @modelcontextprotocol/sdk@1.25.3, @cloudflare/workers-oauth-provider@0.2.2, agents@0.3.6 Last Verified: 2026-01-21 Errors Prevented: 24 documented issues (100% prevention rate) Skill Version: 3.1.0 | Changes: Added IoContext timeout (#23), OAuth remote failures (#24), Security section (PKCE vulnerability), Architecture clarification (internal WebSocket), Tool return format pattern, Conditional tool registration