What it does

code-review-graph parses your repo with Tree-sitter, stores functions, classes, imports and call relationships in a local SQLite graph, then feeds precise context to AI coding tools via MCP. When you change a file, it traces the “blast radius” — callers, dependents, tests — and returns only those files instead of the whole project.

The interesting bit

The benchmarks are unusually honest. The authors openly report 100% recall but only 0.71 average F1 because they deliberately over-predict rather than miss dependencies. They also flag that trivial single-file changes can actually cost more tokens than naive reads, since the graph carries structural metadata overhead. That kind of self-awareness is rare.

Key highlights

• Token reductions from 38x to 528x on real repos (FastAPI, Flask, httpx, etc.), with reproducible pinned-SHA benchmarks
• Incremental re-indexing under 2 seconds for a 2,900-file project via SHA-256 diffing
• Auto-configures MCP for 12+ platforms: Codex, Claude Code, Cursor, Windsurf, Zed, Continue, Copilot, Gemini CLI, Kiro, and others
• 25+ languages including Jupyter notebooks, with Tree-sitter fallbacks where needed
• Local-first: SQLite graph, no cloud dependency, deterministic CPU embeddings

Caveats

• Search quality is middling (MRR 0.35); Express-style module patterns return zero hits currently
• Flow detection only works reliably for Python frameworks; JavaScript and Go entry-point detection is spotty
• For tiny single-file diffs, the graph response can exceed raw file content in token count

Verdict

Worth trying if you’re burning context-window on large repos or monorepos with Claude/Cursor/Copilot. Skip it if your projects are small enough that the AI already sees everything, or if you mostly work in JS/Go and need precise entry-point tracing today.