Deploy MCP Server to Cloudflare Workers: MCP at the Edge

Cloudflare Workers is the ideal edge platform to deploy MCP Server on - low latency, pay-per-request, fast cold start. This article walks the full path.

1. Why Workers

Common options for deploying MCP Server:

Workers' edge: global latency + pay-per-request + KV/D1/R2 adjacent.

2. stdio vs Streamable HTTP

MCP has two transports:

• stdio: local process (Host spawns the server binary, communicates via stdin/stdout)
• Streamable HTTP: HTTP + SSE (Server exposes HTTP endpoint, Host calls via POST + GET-SSE)

Workers can only run Streamable HTTP - Workers is edge runtime, no stdin/stdout processes.

To convert, use @modelcontextprotocol/sdk's StreamableHTTPServerTransport instead of StdioServerTransport:

import { StreamableHTTPServerTransport } from "@modelcontextprotocol/sdk/server/streamableHttp.js";

const transport = new StreamableHTTPServerTransport({
  sessionIdGenerator: undefined,  // stateless
});

server.connect(transport);

// Workers fetch handler
export default {
  async fetch(request: Request, env: Env, ctx: ExecutionContext): Promise<Response> {
    return transport.handleRequest(request);
  }
};

3. Edge Runtime limits

Workers Edge Runtime is not Node.js - the following are unavailable:

• fs (filesystem)
• child_process (subprocess)
• setImmediate / process.nextTick
• any Node.js built-in modules (except Web standards: fetch / crypto / streams / URL)

Code must:

• use only Web standard APIs (fetch, crypto.subtle, TransformStream)
• use Workers KV / D1 / R2 instead of local storage
• long-running tasks via ctx.waitUntil (async)
• cannot require Node packages; use esbuild-compatible packages (most modern SDKs work)

4. CPU time limits

MCP tool calls are usually low CPU intensity (IO-bound) - 5ms suffices in most cases; complex compute (scraping + parsing) goes to Unbound.

5. KV persistence

Workers KV (edge key-value store) fits MCP scenarios:

// Read
const counter = await env.MCP_KV.get("rate_limit:user_123", { type: "json" });

// Write
await env.MCP_KV.put("rate_limit:user_123", JSON.stringify({ count: counter + 1 }));

KV properties:

• Eventual consistency (global convergence < 60s)
• Per-key value ≤ 25MB
• Suits rate limits / caches / prefs

Not for strong consistency (use D1 SQLite for that).

6. wrangler.jsonc config

{
  "name": "my-mcp-server",
  "main": "src/index.ts",
  "compatibility_date": "2026-08-01",
  "compatibility_flags": ["nodejs_compat"],  // compatibility for some Node.js APIs
  "vars": {
    "LOG_LEVEL": "info"
  },
  "kv_namespaces": [
    { "binding": "MCP_KV", "id": "xxx" }
  ],
  "secrets": [
    "GITHUB_CLIENT_SECRET",  // set via wrangler secret put
    "API_KEY"
  ]
}

secrets don't go into git - use wrangler secret put GITHUB_CLIENT_SECRET to configure.

7. Deploy + monitor

# Local dev
wrangler dev

# Deploy
wrangler deploy

# Tail logs
wrangler tail

# Set secrets
wrangler secret put GITHUB_CLIENT_SECRET

Cloudflare dashboard provides:

• Request / error rate real-time monitoring
• CPU time usage (spot performance issues)
• KV read/write volume

Set alerts: 5xx rate > 5% sustained 5 minutes → trigger.

8. Practical gotchas

• nodejs_compat flag: some Node.js packages (especially zod / pnpm packages) need this flag to run
• Cold start: first request 50-100ms warm-up; the user's first MCP call is a bit slower
• Worker size limit: 1MB compressed; with SDK it's typically 200-400KB
• CORS: MCP Host (ChatGPT web) calls your Worker from a different origin; configure CORS to allow * or ChatGPT's domain

9. What's Next

• Build Your Own MCP Server: From Zero to Published - full stdio MCP path
• Model Context Protocol: How MCP Works and How to Build on It - protocol deep dive
• OpenAI API Error Handling and Retry - production stability

Update log

• 2026-08-08: Initial publish