Better Code Review Graph

What better-code-review-graph does

better-code-review-graph turns a repository into a queryable code knowledge graph. Parsing is performed with Tree-sitter, with graph data centered on functions, classes, imports, call relationships, and other structural connections. The server exposes this graph through MCP and can also be used through its crg command-line interface.

The review workflow is designed to reduce the amount of source an agent must inspect. Review context can identify changed files from Git, summarize affected structure, find impacted nodes, include selected source snippets, and return review guidance. The graph also supports semantic search, relationship queries, call-site inspection, inheritance and import analysis, and comparisons between commits.

How it works

Build the graph with the graph tool, either from scratch or incrementally. Other graph actions report statistics, create embeddings, export or import graph data, and produce summaries. The query tool then searches the graph or examines relationships and change impact. It can also identify large functions, sample call-site snippets, detect symbols whose callsite lines moved, and compare nodes added, removed, or modified between two commit SHAs.

The review tool provides two review-oriented paths. context derives a focused review package from a Git diff, while delta compares two commits and can optionally report qualified names whose starting lines changed. Graph nodes and edges include temporal columns for tracking validity across revisions.

Semantic search uses the local ONNX model registry from fastretrieval by default, so the local path requires no API key. An optional cloud embedding chain is available when a different embedding setup is appropriate.

Setup and configuration

The package is distributed on PyPI as n24q02m-crg, and the primary command is crg. It supports Python 3.13. A one-off graph operation can run through uvx, for example:

uvx --python 3.13 --from n24q02m-crg crg graph build --full-rebuild \
  --repo-root /path/to/repo

Install the package with pip install n24q02m-crg when persistent console scripts are preferred. The optional Semgrep security engine is installed with the security extra.

MCP can run over stdio or self-hosted HTTP. For stdio clients, register the package command with uvx --python 3.13 n24q02m-crg and set MCP_TRANSPORT=stdio. Self-hosted HTTP exposes an MCP endpoint at /mcp; the documented local development mode uses loopback, while always-on deployment uses token authentication. Configuration is stored under CRG_CONFIG_DIR or ~/.crg.

The config tool reports server state, changes the log level, clears embedding caches, and manages model-cell setup. Setup can use host-provided model configuration or be switched to local mode.

Tools and capabilities

• graph: Build, update, inspect, embed, summarize, export, and import the knowledge graph.
• query: Search code and resolve relationships, impact, callers, callees, inheritors, importers, line shifts, and commit deltas.
• review: Generate focused context for changed files or compare two commits for review.
• security: Run heuristic or Semgrep scans, list rules, suppress findings, and emit JSON or SARIF reports.
• config: Inspect configuration, manage logging and embedding caches, and control model setup.
• help: Retrieve detailed documentation for tools, security behavior, and usage recipes.

Limitations and notes

The project is CLI-first; its MCP adapter is a secondary interface. HTTP operation is self-hosted rather than provided through a hosted endpoint. The Semgrep scanner requires the optional security installation. Opening a graph created before the schema change automatically applies the migration and saves a pre-migration database backup, so deployments should account for that database transition.