Git
by idosal
GitMCP turns any GitHub repository into a documentation server an AI assistant can query directly, so it fetches real docs and searches actual code instead of guessing from stale training data.
Add to Claude Code (example repo)
claude mcp add --transport http gitmcp https://gitmcp.io/idosal/git-mcp
About
GitMCP is a free, remote MCP server that wraps a GitHub repo (or GitHub Pages site) and gives an AI client structured access to it: fetching the project's primary documentation (preferring an llms.txt file if the repo has one, falling back to the README), running a targeted search across that documentation instead of dumping the whole thing into context, pulling the content behind a specific link found in the docs, and searching the repository's actual source with GitHub's code search. The dynamic generic endpoint (gitmcp.io/docs) does the same thing on demand for whichever repo you point it at, without needing a dedicated per-repo setup.
Because it's hosted, there's nothing to install or run yourself. You point your MCP client at a URL (gitmcp.io/{owner}/{repo}, or {owner}.gitmcp.io/{repo} for GitHub Pages projects) and it's live. That makes it a fast way to give a coding agent accurate, current context on a library that's obscure, that changed its API recently, or that simply postdates the model's training cutoff.
Key features
- Fetches a repo's actual documentation, preferring llms.txt over a plain README when available
- Semantic search over a repo's docs so an agent can query without loading entire files
- Fetches content behind specific URLs referenced from within the documentation
- Searches a repository's real source code via GitHub's code search
- No install required, works as a hosted remote MCP server per repo or via a generic dynamic endpoint
- Works with any MCP-compatible client: Cursor, Claude, VS Code, Windsurf, Cline, and others
Use cases
- Getting accurate, current usage examples for a library that's too new or niche to be in a model's training data
- Reducing hallucinated API signatures when an agent writes code against a fast-moving dependency
- Searching a large repository's actual source instead of relying on the model's memory of similar code
- Pointing an agent at documentation for a project the model has never seen, via the generic gitmcp.io/docs endpoint
Available tools
fetch_documentation
Retrieves the repository's primary documentation, preferring an llms.txt file over a plain README.
search_documentation
Runs a targeted, semantic search over the repository's documentation.
search_code
Searches the repository's actual source code via GitHub code search.
fetch_url_content
Fetches and processes the content behind a specific URL referenced in the docs.
Frequently asked questions
Do I need to run or host anything for this to work?
No. GitMCP is a hosted remote MCP server; you just point your client's MCP config at the repo-specific URL and it's ready.
Does it only work for repos I set up in advance?
No. Beyond the per-repo gitmcp.io/{owner}/{repo} URLs, there's a generic gitmcp.io/docs endpoint that resolves whichever repository you ask about at query time.