Anthropic's official MCP cookbook: copy, paste, caveat emptor

What it does This repo houses the small set of reference servers maintained by the MCP steering group. Each server is a self-contained demo showing how to expose a specific capability—filesystem access, Git manipulation, web fetching, a knowledge-graph memory store—to any LLM client that speaks the Model Context Protocol. You run them via npx or uvx, then wire them into a client like Claude Desktop with a JSON snippet.

The interesting bit The project is explicitly not a production platform. The README slaps a big warning on every server: these are educational sketches, not hardened tools. That honesty is refreshing, and it makes the repo more useful as a teaching device than most “reference architectures” that pretend otherwise.

Key highlights

• Seven active reference servers: Everything, Fetch, Filesystem, Git, Memory, Sequential Thinking, Time
• Fourteen archived servers (GitHub, Postgres, Slack, etc.) moved to servers-archived as maintenance shifted to vendors or the community
• TypeScript servers run with npx -y @modelcontextprotocol/server-<name>; Python servers use uvx or pip
• Client configuration is plain JSON: command, args, and optional env vars
• Links out to ten language-specific MCP SDKs (C#, Go, Java, Kotlin, PHP, Python, Ruby, Rust, Swift, TypeScript)

Caveats

• The README repeats twice that these are not production-ready; security hardening is your job
• Several useful servers (GitHub, SQLite, Puppeteer) have been archived, so check servers-archived if you want them

Verdict Grab this if you’re building an MCP server and want working, minimal examples to crib from. Skip it if you need drop-in, battle-tested integrations—head to the MCP Registry for community servers with real maintenance instead.