pastcode

A minimal MCP server for git history, git blame, and debugging the moment a bug was introduced.

AI coding agents can read the current snapshot of a repo, but they usually cannot answer the question that matters most when debugging: when did this bug get introduced, who changed it, and what did the file look like before the breakage?

pastcode gives any MCP-compatible AI agent read access to your git history so it can investigate bugs the way a senior engineer would: by walking back through time, checking blame, diffing commits, and narrowing the culprit with git bisect.

Quick start

Add a server entry like this to your MCP client configuration:

Claude Desktop

{
  "mcpServers": {
    "pastcode": {
      "command": "npx",
      "args": ["-y", "pastcode"]
    }
  }
}

This works the same way in Claude Code, Cursor, Windsurf, or any other MCP-compatible client that launches stdio-based MCP servers. There are no environment variables to set up, no repo config to install, and no global state to manage. The repo path is passed per tool call, so one running server can work across unlimited repositories.

Not published yet, or testing locally? Use the local path variant above instead.

{
  "mcpServers": {
    "pastcode": {
      "command": "node",
      "args": ["/absolute/path/to/pastcode/index.js"]
    }
  }
}

Related MCP server: gitlog-mcp

Available tools

How agents actually use this

1) “When did this bug get introduced?”

A user asks: “The login form breaks for valid email addresses after the last deploy. When did this start?”

The AI may call:

• list_commits_for_file on the file that owns email validation

• blame_line_range on the suspicious lines

• get_file_at_commit on historical revisions to inspect behavior before and after the regression

• find_introducing_commit with a minimal test command to confirm the exact first bad commit

2) “Who changed this logic and why?”

A user asks: “I want to know why the pricing calculation changed in the checkout flow.”

The AI may call:

• blame_line_range on the calculation block

• diff_between_commits between the relevant revisions

• get_file_at_commit to inspect the code as it existed before the refactor or hotfix

3) “What did this file look like before the refactor?”

A user asks: “The refactor in the auth service looks risky. Show me the pre-refactor implementation.”

The AI may call:

• list_commits_for_file to locate the refactor commit

• get_file_at_commit at the parent ref or historical tag

• diff_between_commits between the refactor and its parent to highlight the exact behavioral change

This is not a CLI tool you run manually. The AI agent drives it as part of a normal debugging workflow.

Security note

⚠️ find_introducing_commit executes an arbitrary shell command (testCommand) during bisect. Only use it against trusted repositories and trusted test commands, just as you would with any AI-driven script execution or automated debugging workflow.

Contributing

This project is intentionally small and dependency-light. The goal is to stay easy to reason about and easy to wire into AI tools without over-engineering the stack.

To work on it locally:

git clone https://github.com/muhcen/pastcode.git
cd pastcode
npm install

For manual inspection of the MCP server, you can run it via the official MCP inspector:

npx @modelcontextprotocol/inspector node index.js

Then connect the inspector to the local server entry and inspect tool calls in real time.

If you want to propose a new tool, open an issue first to discuss scope and fit before sending a PR. This project keeps the surface area intentionally small, and it’s best to align on the workflow before adding new capabilities.

License

This project is licensed under the MIT License.