On a Go and Angular monorepo of 566 nodes and 620 edges, the same bug fix took 29,838 tokens with repo-graph against 75,308 without, exploring 2 files instead of about 15. The README is equally direct that on a small single-language project the model can just read the code; this earns its keep on large repos, monorepos and multi-service systems.
A Python server that scans a repository once and builds a graph of entities, relationships and end-to-end flows, then exposes 6 tools the model uses to find the minimum set of files it actually needs.
- `orient` — the first call on a repo: node and edge counts, detected kinds, entry points, and a blind-spots note flagging which languages or edges are under-linked so you know where to grep instead
- `find` — turns a symbol, a pasted stacktrace, a failing-test id or a diff into the ranked nodes it implicates, every row carrying `path:line`
- `impact` — blast radius forward or backward as a ranked closure, each row with the edge reason and a marker where the engine finds a node unreachable
- `trace` — one feature end to end with each hop labelled by mechanism (call, HTTP, queue, event) and cross-service hops marked; with two arguments, the shortest path between them
- `read` — a node's exact source sliced by its line span, plus a context footer naming HTTP method, cross-stack callers, covering tests and governing docs
- `refresh` — incremental rebuild in which only changed files reparse; routine edits are picked up by the file watcher
Run with `uvx mcp-repo-graph --repo /path/to/your/project`, or `pip install mcp-repo-graph` for a persistent command on your PATH. `--repo` also accepts a public git URL, which it shallow-clones for you (needs `git`). `uvx mcp-repo-graph install` wires up every agent it finds in one command.
One command — uvx mcp-repo-graph install