Why you'd want it
code-review-graph parses a repository with Tree-sitter into a graph of functions, classes, imports, calls, inheritance and test coverage, stored in one local SQLite file. It serves that graph over MCP and a CLI, and its install command detects the AI coding tools you have and writes an MCP server entry for each. Blast-radius analysis traces the callers, dependents and tests a changed file could affect, and a detect-changes tool maps diffs to risk-scored functions, flows and test gaps. Hooks, a pre-commit hook and watch mode update the graph incrementally, re-parsing only files whose hash changed. The README reports a median per-question token reduction of about 63x across 6 repositories against a whole-corpus baseline, and itself calls that baseline an upper bound no real agent pays.
Best for
- Blast-radius context for code review
- A persistent repo map for MCP clients
- Risk-scored PR comments in CI
Use it when
You use a supported AI coding tool such as Claude Code, Cursor or Codex on a sizeable repository and want it to query a structural map instead of re-reading files.
Skip it if
The README's FAQ lists small repos, trivial single-file diffs and one-off questions as poor fits, and its limitations section says graph context can exceed a plain file read for trivial edits, search ranking needs work, JavaScript and Go flow detection needs work, and impact analysis flags false positives in large dependency graphs. The install command writes MCP entries, hooks, skills and rules-file instructions into your tools' configs (an uninstall command is provided). Optional cloud embeddings take API keys and send source-derived text to the provider, though the README says the tool has no telemetry.
Editor's note
The token-reduction figures are the README's own benchmark claims (snapshot of 2026-08-02, see docs/REPRODUCING.md); we have not reproduced them. The README does not label the project alpha or beta; its beta and preview wording is advice about embedding model IDs. Requires Python 3.10+.