Coding agents burn most of their tokens reading files they don’t need — Claude Code eating 33k tokens on startup was a whole discourse. code-review-graph, now the #1 AI repo on GitHub Trending with 20,125 stars and +356 in a day, goes straight at that. It’s a local-first MCP server plus CLI: Tree-sitter parses your repo into a persistent graph — functions, classes, calls, inheritance — in local SQLite. Instead of dumping whole files into context, the agent queries the graph for a change’s “blast radius” and reads only that.
The numbers
Across 6 real repos, median token reduction is ~82x per question, peaking at 528x on FastAPI’s 1,122-file codebase. Incremental updates finish in under 2 seconds. MIT-licensed, 20+ languages, everything stays on your machine.
Plugging it into your agent
pip install code-review-graph, then one install command auto-configures Claude Code, Cursor, Codex, Windsurf, Zed and more. It exposes 30 MCP tools — impact-radius analysis, semantic search, risk-scored change detection. Typical use case: code review on a large monorepo without the agent reading 500 files to answer one question.
Context is the cost center of coding agents right now. This is the picks-and-shovels answer: an index layer for the whole MCP ecosystem.
You Might Also Like
- 27k Github Stars in Weeks Learn Claude Code by Shareai lab Breaks Down ai Coding Agents Into 12 Lessons
- From 739k to 15k Tokens how Code Review Graph Slashes Claude Code Costs With a Local Knowledge Graph
- Claude hud hit 5 3k Github Stars Because Developers Were Flying Blind With Claude Code
- 27 Agents 109 Skills 88k Github Stars is Everything Claude Code Genius or Over Engineering
- From Claude Flow to Ruflo 22k Stars 5900 Commits and the Multi Agent Swarm Taking Over Claude Code
Top AI Product 
Leave a comment