Git MCP is Anthropic’s reference server for letting an agent operate on a local Git repository. It exposes branches, commits, diffs, blame, logs, stashes, and basic write operations (commit, branch, checkout) as structured tools — so the model can answer “what changed in the last week and who did it?” without you running a single git log by hand.
What it produces: a dozen Git tools — git_log, git_diff, git_blame, git_show, git_status, plus write ops (git_commit, git_branch, git_checkout, git_stash). All scoped to a --repository path you set at launch.
Best for: debugging a regression (“git blame the line that broke this test”), code-history archaeology (“when did this function change last and why?”), structured commits inside an agent loop. Pairs perfectly with Claude Code’s diff-first workflow.
Skip if: you’re using GitHub MCP Server — that one already covers most local-repo intent for repos hosted on GitHub. Git MCP shines when you want pure-local, no-API-token operations on a private repo.
Setup gotchas: install uses uvx mcp-server-git (Python-based, requires uv). Easier than npx for many users; harder if your machine doesn’t have Python tooling. Repository path is a launch arg, not a tool param — relaunch the server to switch repos. Don’t mix Git MCP and Filesystem MCP write ops on the same repo without a workflow rule, or you’ll race-condition file edits with stashes.
Real-world workflow: “what did I ship this week?” — agent calls git_log --since='last monday', summarizes the commits grouped by feature area. Replaces my Friday status update generation. Two minutes of agent time.
Compatible alternatives: GitHub MCP Server for hosted-repo workflows, Filesystem MCP for raw file ops.
Use it for local archaeology. Use GitHub MCP for everything that touches the platform.