tirth8205/code-review-graph
code-review-graph gives coding agents a local, incrementally updated dependency map over MCP, narrowing review context while exposing affected callers, dependents, and tests.
code-review-graph parses repositories with Tree-sitter into a persistent local graph, then serves targeted context through MCP. Its six-repository evaluation reports a **~65x median token reduction**, with graph queries typically returning **2,000–3,500 tokens**.
Use it when reviews or impact analysis repeatedly make agents rediscover a large codebase. Hooks and watch mode update only changed files and their dependents; a two-file edit in a roughly **3,000-file repository took about 2.5 seconds** on the measured hook path.
code-review-graph parses repositories with Tree-sitter into a persistent local graph, then serves targeted context through MCP. Its six-repository evaluation reports a **~65x median token reduction**, with graph queries typically returning **2,000–3,500 tokens**. Use it when reviews or impact analysis repeatedly make agents rediscover a large codebase. Hooks and watch mode update only changed files and their dependents; a two-file edit in a roughly **3,000-file repository took about 2.5 seconds** on the measured hook path. The headline comparison uses whole-corpus reading as an upper bound, not a realistic agent workflow. Reported recall is circular because its ground truth comes from the same graph, while search MRR is **0.35** and flow-detection recall is 33%.
This turns graph-shaped code context from a general retrieval pattern into a persistent, incrementally updated MCP implementation for review and impact analysis. It reinforces bounded evidence retrieval over whole-repository loading, while sharply narrowing the performance claim: the token baseline is unrealistic, graph-derived recall is circular, and weak search and flow results leave practical retrieval quality unresolved.