Most AI coding workflows re-send large parts of a repository to the model for every review or task. code-review-graph reverses that pattern by maintaining a persistent, structural map of your code so an assistant only ingests the files and symbols that actually matter for a given change or question.
What Sets It Apart
- Token-first optimization: Produces compact, token-efficient review contexts (benchmarked as ~6.8× fewer tokens for reviews and up to 49× on some daily coding workflows) so LLM-based assistants cost and latency drop in proportion to token savings.
- Structural, incremental indexing: Parses repositories with Tree-sitter into nodes (functions, classes, imports) and edges (calls, inheritance, tests), stores the graph in a local SQLite backend, and updates incrementally on file saves or commits—typical re-index under 2s for large repos.
- Blast-radius & impact analysis: Traces dependents, callers, and tests to compute a conservative "blast radius" for changes so the assistant reads a minimal, high-recall subset instead of the whole codebase.
- MCP-native integration: Exposes ~28 MCP tools (get_minimal_context, detect_changes, query_graph, etc.) so editors and assistants (e.g., Claude Code and other MCP-capable platforms) receive compact, structured context automatically.
- Privacy-first and local-first: No cloud database required—embeddings and optional vector stores are add-ons; core graph lives locally under .code-review-graph.
Who It's For & Tradeoffs
Great fit if you:
- Use LLM-based code review or AI coding assistants frequently and want to cut token costs and latency.
- Work with medium-to-large repos or monorepos where full-context reads are wasteful.
- Need local-first tooling and MCP integration for editor/assistant workflows.
Look elsewhere if you:
- Primarily make tiny, single-file edits (graph metadata can outweigh benefits for trivial changes).
- Rely on cloud-hosted platform features that expect full-file body embeddings by default—CRG focuses on signatures/structural nodes first.
- Need perfect flow detection in all languages; current flow-detection and search ranking have known precision/recall trade-offs (conservative blast-radius, deliberate over-approximation).
Practical notes
Requires Python 3.10+, Tree-sitter grammars for target languages, and optional embedding providers for semantic search. Good for teams wanting deterministic, auditable context bounding for AI assistants; expect an initial build step and light developer config (MCP/CLI) to connect to your tooling.
