code-review-graph

Install code-review-graph, run code-review-graph build in your project, and connect it via MCP. The AI then queries the dependency graph for only the files in a change's blast radius instead of reading everything, reducing tokens by 6.8x on average.

Blast radius analysis uses BFS traversal over a graph of calls, imports, and test coverage edges to identify every file impacted by a change. code-review-graph computes this from Tree-sitter ASTs stored in SQLite and serves results through MCP.

code-review-graph uses Tree-sitter static analysis, is fast, supports 19 languages, and runs locally with no heavy setup. Serena is LSP-based with deeper semantic and type resolution, making it better for polymorphism-heavy or reflection-heavy codebases.

Yes, it supports TypeScript and multi-repo registration for microservice architectures. On a 27,000-file Next.js monorepo it reduced context from 739K to 15K tokens, though path aliases may require tsconfig resolution configuration.

Skip it for codebases under roughly 200 files with isolated single-file changes, where graph overhead exceeds savings. It also cannot trace dynamic imports, reflection-based calls, or runtime-generated code, so those dependencies will be missed.

The graph becomes stale when files change without re-indexing. Run code-review-graph watch during development for incremental updates under 2 seconds, or run code-review-graph update manually before each task.