🔥 Hot Repo: Claude Code Cuts Token Bills 49x in Monorepos

One-liner — code-review-graph builds a persistent structural map of your codebase so Claude Code reads only the files that actually matter, slashing token usage 8.2x on average and up to 49x in large monorepos. - Repo: tirth8205/code-review-graphhttps://github.com/tirth8205/code-review-graph - Stars: ⭐ 14,784 +323 today - Language: Python - License: MIT --- What It Does Claude Code and other AI coding assistants scan your entire codebase on every request — an expensive, slow pattern that scales badly. code-review-graph fixes this by parsing your repo with Tree-sitter into a SQLite-backed graph of functions, classes, imports, and call edges, then exposing that graph via an MCP server. When Claude asks what files matter for a given change, it gets a precise blast-radius answer instead of reading everything. Why It's Blowing Up Token costs for AI-assisted coding compound fast. A single PR review in a mid-size Python repo can run 40–50K tokens with naive full-file reading; this tool cuts that to under 5K by computing the minimal "blast radius" of changed files. Benchmarks across 6 real open-source repos fastapi, flask, Next.js, gin, httpx, express show an 8.2x average reduction, with flask hitting 9.1x, gin at 16.4x, and a Next.js monorepo — 27,700+ files — shrinking review context down to 15 files, a 49x reduction. Every benchmark is reproducible via code-review-graph eval --all. The April v2.3.2 release added 6 new MCP tools hub detection, bridge nodes, knowledge gap analysis, multi-format graph export GraphML, Neo4j Cypher, Obsidian vault, and 4 new languages, pushing support to 23 languages plus Jupyter notebooks. The tool now auto-detects and configures every major AI coding platform — Claude Code, Codex, Cursor, Windsurf, Zed — with a single install command. Key Features - Blast-radius analysis — traces every caller, dependent, and test affected by a change; 100% recall on impact detection in benchmarks - Incremental updates — re-indexes only changed