Local knowledge graph for Claude Code. Builds a persistent map of your codebase so Claude reads only what matters — 6.8× fewer tokens on reviews and up to 49× on daily coding tasks.
Created 2026-02-26 — only ~2.5 months old at appraisal — yet has reached v2.3.3, implying meaningful semantic versioning across at least two minor-version cycles. 44 open issues is well within normal range for a project with 16k stars. The presence of a dedicated eval runner (code-review-graph eval --all) with raw data in evaluate/reports/summary.md and automated CI signals engineering discipline beyond a prototype. One caution: the project has not yet had a full year of production use to surface edge cases, so "stable beta" is the honest classification rather than GA stable.
Exceptional. The README provides a five-language landing page, a three-command quick start, eight architecture diagrams, a reproducible benchmark table with caveats honestly disclosed (the express <1x case is explained), and a --platform-scoped install reference covering 12+ AI coding tools. Beyond the README, a Discord community and dedicated website (code-review-graph.com) exist. The benchmark reproduce path (code-review-graph eval --all) and raw data file give institutional trust to the claimed numbers.
Python 3.10+ requirement is modern and appropriate. Tree-sitter is the correct, battle-tested choice for multi-language AST parsing. SQLite as the graph store is a pragmatic, zero-infra decision that fits the local-first design intent. MCP compatibility indicates adherence to an emerging but actively governed standard. The dependency manifest was unavailable for this appraisal, so dependency hygiene cannot be directly verified — hence 4 rather than 5. CI badge is present and passing per the README shield.
Last commit 2026-05-08, three days before appraisal. Release v2.3.3 published same day as the last commit. Commit cadence over the 2.5-month lifespan has produced a minor-version bump (2.x) plus three patch releases, suggesting active triage and shipping. Issue count (44) relative to star count (16k) is low, implying either fast closure or active moderation. Someone is clearly home.
16,096 stars accrued in ~10 weeks is an extraordinary trajectory — comparable to breakout developer-tool projects. 1,760 forks represents a 10.9% fork ratio, healthy for a tool repo (higher than typical library repos, indicating active forks for customization). A Discord server, multi-language README, PyPI package, and dedicated website all point to an intentional community-building effort, not organic accretion alone.
Overall: 4.6/5
Category: Code Knowledge Graph Known alternatives in vault: None — this is the first repo appraised in this category. Differentiation: code-review-graph occupies a specific niche: persistent, incremental, AST-level codebase indexing surfaced via MCP to AI coding assistants, with explicit blast-radius computation. Generic context-stuffing tools (e.g., repomix, gpt-repository-loader) dump entire repos into prompts without structural pruning. Aider and similar agents do lightweight context selection but lack a persistent graph or incremental SHA-256 diffing. The monorepo funnel benchmark (27,732 → ~15 files) is a concrete differentiation claim with reproducible methodology. The 24-language + Jupyter coverage is broader than most comparable tools. Gap or crowd: Clear gap in the vault. No competing appraisal exists; this is a founding entry in the Code Knowledge Graph category.
Score: 5/5
Harvestable: Blast-radius graph traversal algorithm (caller/dependent/test tracing); SHA-256 incremental diff pattern for re-indexing only changed nodes; MCP server wiring pattern reusable for any Python tool exposing structured context to Claude; Tree-sitter multi-language AST extraction scaffolding for 24 languages including Jupyter notebooks.
Integration path: Direct — install via pip install code-review-graph, run code-review-graph install, and the tool auto-injects itself as an MCP server into Claude Code or any supported AI coding platform. No wrapper needed. For a PAI system with a skills/hooks layer, the code-review-graph build and incremental git hooks map cleanly onto existing build-on-commit automation patterns.
Overlap with existing: No vault overlap detected. If the PAI system already implements a custom repo-indexer or context-selector, there is potential functional duplication, but this repo's MCP surface and blast-radius logic would still be worth harvesting as a reference implementation.
Adoption cost: Trivial — single pip install, single CLI command, MCP config auto-generated. Deeper integration (embedding the graph engine as a library within a PAI skill) would be moderate effort, but out-of-the-box value is immediate.
The token-reduction claims (8.2× average, 49× on Next.js monorepo) are unusually well-substantiated for a new project: the benchmark methodology names six specific open-source repos, specifies commit counts, provides raw per-repo numbers, acknowledges the express anomaly, and exposes a reproduce command. This level of epistemic honesty raises confidence in the numbers. The rapid star growth (16k in 10 weeks) suggests word-of-mouth in the Claude Code and MCP communities, which are currently high-velocity ecosystems. The multi-platform support (Codex, Cursor, Windsurf, Zed, Gemini CLI, GitHub Copilot, Kiro, etc.) significantly broadens the addressable user base beyond the Claude-only origin. One watch item: dependency manifest unavailability means supply-chain hygiene is unverified — worth checking pyproject.toml / requirements.txt before production integration.