better-code-review-graph

Knowledge graph for token-efficient code reviews with Tree-sitter parsing, dual-mode embedding (ONNX + LiteLLM), and blast-radius analysis via MCP tools.

GitHub

better-code-review-graph

mcp-name: io.github.n24q02m/better-code-review-graph

Knowledge graph for token-efficient code reviews -- fixed search, configurable embeddings, qualified call resolution.

CI codecov PyPI Docker License: MIT

Python MCP semantic-release Renovate

Fork of code-review-graph with critical bug fixes, configurable embeddings, and production CI/CD. Parses your codebase with Tree-sitter, builds a structural graph of functions/classes/imports, and gives Claude (or any MCP client) precise context so it reads only what matters.

Why Better

Featurecode-review-graphbetter-code-review-graph
Multi-word searchBroken (literal substring match)AND-logic word splitting ("firebase auth" matches both verify_firebase_token and FirebaseAuth)
callers_of accuracyEmpty results (bare name targets)Qualified name resolution -- same-file calls resolved to file::name
Embedding modelall-MiniLM-L6-v2 + torch (1.1 GB)qwen3-embed ONNX + LiteLLM (200 MB)
Output sizeUnbounded (500K+ chars possible)Paginated (default 500 nodes, truncation metadata)
Plugin hooksInvalid PostEdit/PostGit eventsValid PostToolUse (Write, Edit, Bash)
Plugin MCPDuplicate registration (.mcp.json + plugin.json)Single source (plugin.json only)
Python version3.10+3.13 (pinned)
CI/CDGitHub Actions basicPSR + Docker multi-arch + MCP Registry
Test coverageUnknown95%+ enforced

All fixes are submitted upstream as standalone PRs (see Upstream PRs). If all are merged, this repo will be archived.

Quick Start

Prerequisites

  • Python 3.13 (required -- requires-python = "==3.13.*")

Option 1: uvx (Recommended)

{
  "mcpServers": {
    "better-code-review-graph": {
      "command": "uvx",
      "args": ["--python", "3.13", "better-code-review-graph", "serve"],
      "env": {
        // -- optional: cloud embeddings via LiteLLM
        // "API_KEYS": "GOOGLE_API_KEY:AIza...",
        // -- optional: LiteLLM Proxy (selfhosted gateway)
        // "LITELLM_PROXY_URL": "http://10.0.0.20:4000",
        // "LITELLM_PROXY_KEY": "sk-your-virtual-key"
        // -- without API_KEYS, uses built-in local qwen3-embed ONNX (zero-config)
      }
    }
  }
}

Option 2: pip

pip install better-code-review-graph
better-code-review-graph install   # creates .mcp.json in project root

Option 3: Docker

{
  "mcpServers": {
    "better-code-review-graph": {
      "command": "docker",
      "args": [
        "run", "-i", "--rm",
        "-v", "crg-data:/data",
        "-e", "API_KEYS",
        "n24q02m/better-code-review-graph:latest"
      ],
      "env": {
        // -- optional: cloud embeddings
        // "API_KEYS": "GOOGLE_API_KEY:AIza..."
      }
    }
  }
}

Option 4: Claude Code Plugin

claude plugin install n24q02m/better-code-review-graph@better-code-review-graph

Then open your project and tell Claude:

Build the code review graph for this project

MCP Tools

Claude uses these automatically once the graph is built.

ToolDescription
build_or_update_graph_toolBuild or incrementally update the graph. Default: incremental (changed files only).
get_impact_radius_toolBlast radius of changed files. Shows which functions, classes, files are affected. Paginated with max_results.
get_review_context_toolToken-optimized review context with structural summary, source snippets, and review guidance.
query_graph_toolPredefined queries: callers_of, callees_of, imports_of, importers_of, children_of, tests_for, inheritors_of, file_summary.
semantic_search_nodes_toolSearch code entities by name/keyword or semantic similarity (requires embeddings).
embed_graph_toolCompute vector embeddings for semantic search. Uses dual-mode backend.
list_graph_stats_toolGraph size, languages, node/edge breakdown, embedding count.
get_docs_section_toolRetrieve specific documentation sections for minimal token usage.
find_large_functions_toolFind functions/classes exceeding a line-count threshold for decomposition audits.

Embedding Backends

Embeddings enable semantic search (vector similarity instead of keyword matching). Two backends are available:

BackendConfigSizeDescription
local (default)Nothing needed~570 MB (first use)qwen3-embed ONNX. Zero-config. Downloaded on first embed_graph_tool call.
litellmAPI_KEYS or LITELLM_PROXY_URL0 MBCloud providers via LiteLLM (Gemini, OpenAI, Cohere, etc.).
  • Auto-detection: If API_KEYS or LITELLM_PROXY_URL is set, uses LiteLLM. Otherwise, uses local ONNX.
  • Override: Set EMBEDDING_BACKEND=local or EMBEDDING_BACKEND=litellm explicitly.
  • Fixed 768-dim storage: All embeddings stored at 768 dimensions via MRL truncation. Switching backends does NOT invalidate existing vectors.
  • Lazy loading: Model downloads on first embed call, not on server start.

CLI Reference

better-code-review-graph install     # Register MCP server with Claude Code (creates .mcp.json)
better-code-review-graph init        # Alias for install
better-code-review-graph build       # Full graph build (parse all files)
better-code-review-graph update      # Incremental update (changed files only)
better-code-review-graph watch       # Auto-update on file changes
better-code-review-graph status      # Show graph statistics
better-code-review-graph serve       # Start MCP server (stdio transport)

Configuration

VariableDefaultDescription
EMBEDDING_BACKEND(auto-detect)local (qwen3-embed ONNX) or litellm (cloud API). Auto: API_KEYS/proxy -> litellm, else local.
EMBEDDING_MODELgemini/gemini-embedding-001LiteLLM embedding model (only used when backend=litellm).
API_KEYS-LLM API keys for SDK mode (format: ENV_VAR:key,...). Enables LiteLLM backend.
LITELLM_PROXY_URL-LiteLLM Proxy URL. Enables LiteLLM backend via proxy.
LITELLM_PROXY_KEY-LiteLLM Proxy virtual key.

Ignore files

Create .code-review-graphignore in your project root to exclude paths:

generated/**
*.generated.ts
vendor/**
node_modules/**

Supported Languages

Python, TypeScript, JavaScript, Go, Rust, Java, C#, Ruby, Kotlin, Swift, PHP, C/C++

Each language has full Tree-sitter grammar support for functions, classes, imports, call sites, inheritance, and test detection.

Cross-Agent Compatibility

FeatureClaude CodeCopilot CLICodexGemini CLIAntigravityOpenCodeCursorWindsurfClineAmp
MCP tools (9 tools)YesYesYesYesYesYesYesYesYesYes
CLAUDE.md / AGENTS.mdYesYesYesYesYesYesYesYes----
Skills (slash commands)YesYesYesYes--Yes--------
Hooks (PostToolUse)Yes--YesYes------------
Plugin (marketplace)YesYes----------------

Upstream PRs

All fixes in this fork are submitted as standalone PRs to the original code-review-graph:

  • #37 -- Multi-word search AND logic
  • #38 -- Parser call target resolution (fixes #20)
  • #39 -- Impact radius output pagination

If all upstream PRs are merged, this repository will be archived.

Build from Source

git clone https://github.com/n24q02m/better-code-review-graph
cd better-code-review-graph
uv sync --group dev
uv run pytest
uv run better-code-review-graph serve

Requirements: Python 3.13 (not 3.14+), uv

Compatible With

Claude Desktop Claude Code Cursor VS Code Copilot Antigravity Gemini CLI OpenAI Codex OpenCode

Also by n24q02m

ServerDescriptionInstall
wet-mcpWeb search, content extraction, library docsuvx --python 3.13 wet-mcp@latest
mnemo-mcpPersistent AI memory with hybrid searchuvx mnemo-mcp@latest
better-notion-mcpNotion API for AI agentsnpx -y @n24q02m/better-notion-mcp@latest
better-email-mcpEmail (IMAP/SMTP) for AI agentsnpx -y @n24q02m/better-email-mcp@latest
better-godot-mcpGodot Engine for AI agentsnpx -y @n24q02m/better-godot-mcp@latest
better-telegram-mcpTelegram Bot API + MTProto for AI agentsuvx --python 3.13 better-telegram-mcp@latest

License

MIT - See LICENSE

관련 서버