CodeGraph

README INTEGRITY ISSUE. The architecture diagrams and the 8.2× token-reduction benchmark table in the upstream README are copied verbatim from tirth8205/code-review-graph. The diagrams/ directory does not exist in this repo (diagram image refs are broken); the eval command cited (code-review-graph eval --all) is CRG’s CLI, not this tool’s. The 94% fewer-tool-calls claim may be genuine — a separate eval runner exists in the test suite (evaluation/runner.ts) — but the remainder of the upstream benchmark section is CRG’s data presented as this tool’s own. Do not rely on the README benchmark figures without independent reproduction.

TL;DR

• Node.js MCP server that pre-indexes a codebase into a Tree-sitter AST-backed SQLite knowledge graph so Claude Code agents query structure instead of scanning files.
• Reports 94% fewer tool calls and 77% faster exploration (as reported, README — provenance uncertain; see integrity notice above).
• 100% local — WASM-bundled tree-sitter and SQLite; no cloud dependency.
• Supports 19 languages + Jupyter notebooks; single MCP tool codegraph_explore.
• npm package @colbymchenry/codegraph v0.7.2; 412 stars.

What’s novel / different

CodeGraph inserts a pre-built structural index between the agent and the file system. Instead of spawning grep/glob/read calls, the Explore agent calls codegraph_explore once and receives blast-radius context: which symbols were touched, what calls them, and what tests cover them. The approach is independent from but architecturally similar to tirth8205/code-review-graph — created first (2026-01-18 vs CRG’s 2026-02-26), but eclipsed by CRG in adoption. Uniquely, uses WASM-bundled tree-sitter (no native deps) making install completely dependency-free. Lean surface: single MCP tool vs CRG’s 22.

Architecture overview

Core design

Repository is parsed by Tree-sitter into nodes (functions, classes, imports) and edges (calls, inheritance, test coverage) stored in SQLite. At query time, codegraph_explore computes the minimal file set needed for the agent’s question via blast-radius traversal.

Interface / API

Exposed as an MCP server. Configured once via claude_mcp_config.json. Key tool: codegraph_explore.

Dependencies

• Runtime: Node.js 18+
• Language: TypeScript
• Parser: Tree-sitter
• Storage: SQLite (local)
• No cloud or external model dependency

Scope / limitations

• 19 languages documented; coverage beyond those is unspecified.
• Benchmarks run on 6 open-source repos with a single Explore-agent query type — may not generalise to other agent patterns.
• Architecture diagrams and the 8.2× benchmark table in the README are copied from tirth8205/code-review-graph; treat upstream README benchmark section as unreliable for this tool.

Deployment model

• Runtime: Node.js 18+, local machine
• Install: npm install -g @colbymchenry/codegraph or npx
• Storage: SQLite file in project root
• MCP config: claude_mcp_config.json

Self-reported metrics

• 94% fewer tool calls vs no graph (as reported, README)
• 77% faster exploration on average across 6 codebases (as reported, README)
• Per-repo range: 84–96% fewer tool calls (as reported, README)

Open questions

• README provenance (resolved): codegraph and CRG are independent tools (neither is a GitHub fork). codegraph was created first (2026-01-18); CRG launched 2026-02-26. The shared README content — architecture diagrams, the 8.2× token-reduction table, and the code-review-graph eval --all command — was copied from CRG into codegraph’s README. The diagrams/ directory does not exist in codegraph’s repo; the diagram image refs in the README are broken.
• The 94% fewer tool calls / 77% faster figures may be genuine: codegraph has its own TypeScript eval runner (evaluation/runner.ts) that tests against .codegraph/codegraph.db. But these figures have not been independently reproduced.
• The 8.2× token-reduction table in codegraph’s README is CRG’s benchmark data, not codegraph’s — treat as unreliable for codegraph specifically.
• No stated MRR or recall metric for codegraph_explore queries.
• Incremental update latency on large monorepos not documented.