🔍 gitlog-mcp

Give your AI coding agent superpowers over git history.

gitlog-mcp is a Model Context Protocol (MCP) server that lets AI agents (Claude Code, Cursor, Windsurf, and any MCP client) understand what changed in a repository — auto-generate changelogs, analyze commits, attribute blame, and draft release notes.

"What changed in this repo?" is the question every AI agent gets wrong. This fixes it.

Why this exists

AI coding agents are great at writing code but famously bad at knowing the history of a codebase. They hallucinate changelogs, misattribute blame, and guess at release notes. gitlog-mcp gives them a reliable, structured window into git log — so their answers are grounded in what actually happened, not invented.

Features

• 📝 Auto-changelog — generate a clean, grouped changelog from any tag/commit range
• 🔎 Commit analysis — explain why a change happened, not just what changed
• 👤 Blame attribution — who touched what, and when (with context)
• 🏷️ Release notes — draft release notes from tag-to-tag diffs
• 📊 Repo health — commit frequency, top contributors, churn hotspots
• 🧱 Zero runtime dependencies — pure Python stdlib + the MCP SDK, one file

Quick start

# 1. Install from PyPI
pip install gitlog-mcp

# 2. Run standalone (for testing) — defaults to the current directory
gitlog-mcp
gitlog-mcp --repo /path/to/your/repo

# 3. Or add directly to your agent's MCP config (see below)

Want the local web dashboard too? It's an optional extra, not installed by default: pip install "gitlog-mcp[ui]". Plain pip install gitlog-mcp stays exactly as dependency-free as ever.

Installing from source (for contributors): git clone https://github.com/ManiaSacha/gitlog-mcp.git && cd gitlog-mcp && pip install -e . — see CONTRIBUTING.md for details.

Claude Code config

{
  "mcpServers": {
    "gitlog": {
      "command": "gitlog-mcp",
      "args": ["--repo", "."]
    }
  }
}

Cursor config (.cursor/mcp.json)

{
  "mcpServers": {
    "gitlog": {
      "command": "gitlog-mcp",
      "args": ["--repo", "."]
    }
  }
}

Windsurf config

{
  "mcpServers": {
    "gitlog": {
      "command": "gitlog-mcp",
      "args": ["--repo", "."]
    }
  }
}

Debug it standalone

Want to poke at the tools directly before wiring up an agent? The MCP Inspector gives you a UI to call each tool by hand:

npx @modelcontextprotocol/inspector gitlog-mcp --repo .

Example agent prompts

"Generate a changelog of everything that changed between v1.2.0 and v1.3.0."

"Who introduced this buggy line in src/parser.py and why?"

"Draft release notes for the next version based on recent commits."

Tools exposed

Web dashboard (optional)

Prefer a browser to a terminal? gitlog-mcp-ui serves a small local dashboard on top of the same git-reading code the MCP tools use — same data, human-readable.

pip install "gitlog-mcp[ui]"
gitlog-mcp-ui --repo /path/to/repo

This prints a URL (http://127.0.0.1:8765 by default) — open it in your browser; it won't open one for you.

Read-only — no write actions anywhere. Local-only by construction: binds to 127.0.0.1 only (there's no --host flag, so it can't be exposed to the network even by accident), and validates every request's Host header too, closing the DNS-rebinding gap a loopback bind alone doesn't cover. No auth needed, because nothing but your own machine can reach it.

It's an optional extra (pip install gitlog-mcp[ui]), not installed by default. The core gitlog-mcp server has zero runtime dependencies beyond the MCP SDK, full stop — the dashboard doesn't change that. It's http.server from the stdlib, no framework, no extra deps of its own.

Architecture

gitlog-mcp (single file, ~300 lines)
├── FastMCP server (stdio transport)
├── GitRunner — thin wrapper over `git` CLI
└── Tools — each maps to a git subcommand + parsing

Deliberately small. You can read the whole thing in an afternoon — that's a feature.

Contributing

PRs welcome. Small, focused, well-tested changes only. See CONTRIBUTING.md.

Release process and versioning are documented in RELEASING.md.

Roadmap

• --repo auto-detection from cwd
• Structured JSON output for all tools
• GitHub/GitLab remote integration
• Tests + CI badge

License

MIT © 2026 — built in the open, for the open-source community.

Star this repo if you want AI agents to stop hallucinating your changelog. ⭐