mcp-quickstart

⭐ 推荐： npx ship-skills init（已登录 gh 会自动支持作者）

Forge a working, testable, publishable MCP server in 30 seconds. One command. TypeScript or Python. Example tools, the Inspector, and tests — all wired up. Or point it at an OpenAPI spec and get one MCP tool per API endpoint, automatically.

npm create mcp-quickstart@latest my-server

Most "create an MCP server" guides leave you with an empty skeleton and a 20-step setup. mcp-quickstart hands you a server that runs, tests, and connects to Claude / Cursor on the first try — then gets out of your way.

Turn any REST API into an MCP server

Most scaffolders hand you an empty template and say "now go write your tools." mcp-quickstart can instead read an existing API and write the tools for you — from an OpenAPI spec or a single curl command:

# whole API, from an OpenAPI spec:
npx mcp-quickstart petstore-mcp --from-openapi https://petstore3.swagger.io/api/v3/openapi.json

# one endpoint, straight from a curl you already have:
npx mcp-quickstart my-tool --from-curl "curl https://api.example.com/v1/search?q=hi -H 'Authorization: Bearer X'"

Point it at any OpenAPI 3.x document (a URL or a local .json / .yaml) and it generates a real, typed MCP server — one tool per operation, with path/query params and request bodies wired up to a small HTTP client. Request bodies are expanded into per-field zod types ($ref, nested objects, arrays and enums resolved), not an opaque blob — so the AI gets a real schema for each tool. Set API_BASE_URL (and an optional auth header) in .env, run npm run dev, and your AI agent can call the API immediately.

--from-curl does the same for a single request: paste a curl command (or a file with one) and get one MCP tool. Query params and the JSON body become overridable tool inputs, and auth headers are moved to .env instead of being hard-coded into the generated source.

my-api-mcp/
├── src/
│   ├── index.ts   # one registerTool(...) per API operation — plain, editable TS
│   └── http.ts    # generic caller: URL building, auth header, body
└── .env.example   # API_BASE_URL, API_AUTH_HEADER, API_AUTH_VALUE

No magic, no runtime spec parsing — just generated TypeScript you own and can edit.

Deploy a remote MCP server to the edge

Local stdio servers are great, but the ecosystem is moving to remote MCP servers over Streamable HTTP. mcp-quickstart can scaffold one that deploys to Cloudflare Workers — stateless (createMcpHandler, no Durable Objects), so it runs on the free Workers plan, cold-starts in milliseconds, and is reachable globally:

npx mcp-quickstart edge-server --transport cloudflare -y
cd edge-server && npm install
npx wrangler login
npm run deploy        # → https://edge-server.<you>.workers.dev/mcp

Then point Cursor / Claude at the /mcp URL. This is exactly what edge runtimes are good at — low-latency orchestration, not running models. (To run a model you still call a GPU backend; the Worker is the global front door / gateway.)

The closed loop: API → globally-deployed remote MCP server

Combine the two ideas. Feed an OpenAPI spec (or a single curl) and target Cloudflare, and you get an API-backed remote MCP server that's ready to deploy worldwide — in one command: