The official reference servers for the Model Context Protocol (MCP), kept by the MCP steering group. The maker says they are reference implementations: educational examples of MCP features and SDK use, not production-ready solutions.
Current servers: Everything (test server with prompts, resources and tools), Fetch (web pages as Markdown), Filesystem (file operations with access controls), Git (read, search and change Git repos), Memory (persistent memory in a local knowledge graph), Sequential Thinking (step-by-step problem solving) and Time (time and timezone conversion). Archived: AWS KB Retrieval, Brave Search, EverArt, GitHub, GitLab, Google Drive, Google Maps, PostgreSQL, Puppeteer, Redis, Sentry, Slack and SQLite.
Each server runs locally and is added to an MCP client such as Claude Desktop. TypeScript servers start with npx, Python servers with uvx or pip, or use Docker.
Free. New code is under the Apache License 2.0; older code stays under MIT during the licence change.
Developers who build their own MCP servers and want working examples of MCP features and the official SDKs, and users of MCP clients such as Claude Desktop who want to connect local files, Git, web pages or memory.
Works on local Git repositories and local files with no GitHub account, and includes Fetch, Memory, Time and Sequential Thinking servers.
The Fetch server is free, runs locally and needs no API key or credits to turn web pages into Markdown.
GitHub's official MCP server that connects AI tools to GitHub
Works with GitHub itself (issues, pull requests, Actions runs, security alerts) and is hosted by GitHub; the reference Git server only reads and changes a local Git repository.
讨论
登录以参与讨论