code-review-graph

NOTE. The README of colbymchenry/codegraph copies this tool’s architecture diagrams, benchmark table, and eval command (code-review-graph eval --all) without attribution. This file is the original source.

TL;DR

• Python MCP server (pip install code-review-graph) that indexes a codebase via Tree-sitter into a SQLite knowledge graph; exposes 22 MCP tools and 5 prompt templates.
• Reports 8.2× average token reduction on reviews (naive vs graph) and up to 49× on daily coding tasks (as reported, README).
• Auto-installs across Claude Code, Cursor, Windsurf, Zed, Continue, OpenCode, and Antigravity via code-review-graph install.
• 19 languages + Jupyter/Databricks notebooks; incremental updates under 2 seconds (as reported).
• Optional semantic search via sentence-transformers, Google Gemini embeddings, or MiniMax.
• Community detection (Leiden algorithm), wiki generation, interactive D3.js visualisation, and multi-repo registry — feature set significantly broader than comparable tools.
• 7,624 GitHub stars (as of 2026-04-08).

What’s novel / different

Among graph-based code intelligence tools, code-review-graph distinguishes itself by combining blast-radius analysis with execution flow tracing, community detection, and LLM-assisted wiki generation — all in a single Python package. The 22-tool MCP surface (vs the 14 tools in deusdata-codebase-memory-mcp or the single codegraph_explore in colbymchenry/codegraph) gives agents fine-grained access to refactoring suggestions, flow criticality ranking, and cross-repo search. The code-review-graph install auto-detection across 8 AI coding platforms is the most multi-platform install story in this category. Architecturally similar to colbymchenry/codegraph (TypeScript) — independent tools, not forks of each other; CRG launched after codegraph but rapidly eclipsed it.

Architecture overview

Core design

Repository parsed by Tree-sitter into SQLite nodes and edges. Blast-radius, dependency chains, and test-coverage gaps are pre-computed. Optional vector embeddings (sentence-transformers, Gemini, MiniMax) stored alongside for hybrid FTS5+vector search. Community detection via Leiden algorithm groups related code; execution flows trace call chains by criticality.

Interface / API

22 MCP tools including: get_impact_radius_tool, get_review_context_tool, query_graph_tool, semantic_search_nodes_tool, detect_changes_tool, refactor_tool, generate_wiki_tool, cross_repo_search_tool. 5 prompt templates: review_changes, architecture_map, debug_issue, onboard_developer, pre_merge_check. Slash commands: /code-review-graph:build-graph, /code-review-graph:review-delta, /code-review-graph:review-pr.

Dependencies

• Runtime: Python 3.10+
• Parser: Tree-sitter
• Storage: SQLite (.code-review-graph/); no external database
• Optional: sentence-transformers (local embeddings), igraph (community detection), ollama/Gemini/MiniMax (embeddings)
• Visualisation: D3.js (HTML output)

Scope / limitations

• MRR 0.35 for keyword search (stated in benchmarks section — low).
• Small single-file changes may produce more tokens than naive file read (express benchmark: 0.7× reduction).
• Semantic embeddings require optional install and external model; disabled by default.
• Community detection requires optional igraph install.

Deployment model

• Runtime: Python 3.10+, local machine
• Install: pip install code-review-graph or pipx install code-review-graph
• Storage: SQLite in .code-review-graph/ directory
• Multi-platform auto-config: code-review-graph install

Self-reported metrics

• 8.2× average token reduction across 6 repos (as reported, README); range 0.7×–16.4×
• Up to 49× token reduction on daily coding tasks (as reported, README — basis not specified)
• Incremental updates under 2 seconds (as reported)
• MRR 0.35 for keyword search (stated, README — low)

Open questions

• Relationship to colbymchenry/codegraph (resolved): independent tools, neither forked from the other. codegraph was created first (2026-01-18); CRG launched 2026-02-26. CRG is the more mature and widely adopted tool. codegraph’s README copies CRG’s diagram assets and benchmark table — the reverse is not true; CRG’s README is the original source of this content.
• “Up to 49× token reduction on daily coding tasks” — no benchmark data or methodology given for this figure.
• No independent benchmark reproduction; code-review-graph eval --all runner (evaluate/) not verified against the reported numbers.
• Community detection quality not benchmarked; Leiden clustering on arbitrary codebases may produce noisy clusters.
• Active development pace (commits co-authored by Claude Opus 4) — review quality of AI-assisted commits for correctness.