code-review-graph Claims ~82x Median Token Reduction — a Tree-Sitter AST Graph in SQLite Exposing 30 MCP Tools, +4,577 Stars This Week

GitHub·high signalXBlueskyLinkedInCopy link

This Python project parses repos into a Tree-sitter AST graph of functions, classes and imports linked by call, inheritance and test-coverage edges, stores it in SQLite under `.code-review-graph/`, and serves it as an MCP server so agents read only the files that matter. Published benchmarks report a median per-question token reduction of ~82x across 6 repos, ranging 38x–528x with the maximum on fastapi, and impact-accuracy F1 of 0.71 (precision 0.44–0.79). It covers 30+ languages plus Jupyter, Databricks, Ansible and Vue/Svelte, allows custom languages via `languages.toml` without forking, and exports to GraphML, Neo4j Cypher and Obsidian.