repomap

repomap vision ​

What we are building ​

repomap gives AI coding agents, and the people who work with them, a map of a codebase: its modules, the files and symbols that matter, how calls flow, and what a change will break. It parses the code itself with tree-sitter, builds one graph, and answers in a few hundred tokens and a few milliseconds.

An agent that is about to change a function should know who calls it, which modules it touches and which tests to run, before it edits, not after review. That is the one job.

Who it is for ​

  • Developers who use an AI coding agent (Claude Code, Codex, Cursor, VS Code, Windsurf, Gemini CLI) on codebases too large to read into a prompt.
  • Teams that want every pull request to show its blast radius, and that work across more than one repository.

Boundaries ​

  • Local first: one binary, no account and no API key. Indexing and queries make no network calls; the only download is the embedding model, once.
  • Read-only: repomap never changes the code it maps. Live database connections stay read-only and never store or print the connection string.
  • Six MCP tools: map, search, context, trace, impact, db. New abilities, Team included, go into those tools, not into more tools.
  • Build-free: answers come from parsing, never from running or building the project. Compiler-grade enrichment (SCIP, language servers) may be added where a build already exists, and the default stays build-free.
  • Hosted only where a team needs it: repomap Team's PR review and shared maps run on Sylphx. The review reads a pull request's code to analyse it and keeps none of it. Shared maps hold paths and symbol names, never source. Everything else runs on the user's machine.

Free and Team ​

Everything repomap does for free stays free, under MIT, including impact in your own CI. repomap Team sells only new value for teams: the hosted PR review, the multi-repository graph, private shared maps, and support. Public repositories get the PR review free. A capability never moves from free to Team. The Team page has the price and what each part does.

What good looks like ​

  • Search leads the published benchmarks (SWE-bench Verified localisation Acc@10, NDCG@10 per language), measured in CI on GitHub-hosted runners.
  • A first map of a large repository takes seconds, a query takes milliseconds, and a map fits the token budget the agent asks for.
  • impact names the callers and tests a change actually reaches, measured against a caller/callee gold set.
  • A pull request review appears within a minute of the push.

Current capabilities and their code: capabilities.md.