AI Native Landscape

code-review-graph

Tracked

A local-first code intelligence graph served over MCP and CLI. It builds a persistent structural map of a codebase with Tree-sitter and updates it incrementally, so AI coding tools read only the files a change actually touches.

Authortirth8205Open Sourced2026-02-26Last CommitUnknown

Overview

code-review-graph builds a local knowledge graph of a codebase so AI coding tools get precise review context instead of re-reading large parts of the repo. It parses code into ASTs with Tree-sitter, stores functions, classes, imports, and their relationships as a graph, keeps the graph updated incrementally through hooks and watch mode, and serves compact context over MCP. Reviewing a change then means querying the blast radius — every caller, dependent, and test that could be affected — and reading only that minimal file set.

Key Features

  • One-command install that detects installed AI coding tools and configures MCP entries, hooks, and skills for each (Claude Code, Codex, Cursor, Windsurf, Zed, Continue, OpenCode, Gemini CLI, Copilot, and more)
  • Blast-radius analysis: when a file changes, traces every caller, dependent, and test that could be affected
  • Incremental updates via hooks and watch mode after an initial build (a ~3,000-file cold build takes about 40 seconds)
  • Benchmarked context reduction — reading flask's whole corpus costs 143,594 tokens where a graph answer costs 2,196 (65x fewer)
  • Clean uninstall that removes only CRG-owned files and entries, leaves other MCP servers and JSONC comments intact, and supports --dry-run preview
  • Partial-build tolerance: files that fail to parse are named in the summary and keep their previous graph rows

Use Cases

  • AI-assisted code review that reads only the files a change touches, not the whole repository
  • Large-repo workflows where re-reading the codebase per prompt blows the context window
  • Serving reusable structural context (call graph, inheritance, test coverage links) to any MCP-capable coding tool
  • GitHub Action-based review pipelines with locally reproducible benchmarks

Technical Details

  • Python 3.10+ package (pip install code-review-graph); repositories are parsed into ASTs with Tree-sitter and persisted as a graph of nodes (functions, classes, imports) and edges (calls, inheritance, test coverage)
  • The graph is queried at review time for the smallest set of files the assistant needs, via MCP tools that return blast radius and risk scores
  • Incremental engine: hooks and watch mode update the graph on file changes; a failed parse yields a partial status that names the files and preserves their previous rows
  • The installer writes per-platform MCP entries choosing the right launcher — poetry run/uv run inside Poetry or uv projects, uvx code-review-graph serve when uvx is on PATH, otherwise the current interpreter
  • Shared config files are replaced atomically so a failed write leaves the original intact; uninstall removes only CRG-owned entries