Skip to content

repo-reader-mcp

A read-only MCP server that lets any AI agent read code in git repositories — any branch, tag or commit, with its history — and never changes them.

Any ref, with its history

Files and folders at any branch, tag or commit; the history of a repository or a path; a commit with its diff; the difference between two refs; who wrote each line, and why.

Search across repositories

One question over every repository at once — text or a regular expression, answered with files and line numbers in a fraction of a second; grep in one repository at any ref, with lines around each match.

Read-only by construction

A fixed set of reading git commands, no shell, an agent’s ref resolved to a commit id before git sees it, no network while reading. Every protection has a test that fails when it is taken out.

Any code host

Local repositories, and remote ones mirrored from GitHub, GitLab, Bitbucket or any git server — one by one, or a whole group by a pattern, new repositories joining as they are created.

Only what you serve

include and exclude hold for every tool: a path left out is never listed, read, searched or diffed — and on a mirror, its contents never reach the server’s disk. Files named for credentials stay out, and values that look like secrets are hidden in every answer.

Guidance for agents

A repository’s AGENTS.md, CLAUDE.md, rules and skills reach the agent reading it — and what the people who run the server add, for one repository or for all.

Any MCP client

stdio for one person’s agent; Streamable HTTP behind a token for a whole team — Claude, Cursor, Codex, any client that speaks MCP.

Status: pre-alpha. It runs from a clone of the source. A published package comes next — the plan.