Dab's Evaluation: This is an extremely valuable tool for our daily operations. It maps code directories into a local SQLite-backed knowledge graph of symbol references, call hierarchies, and web routes. AI agents (like Claude Code, Codex, or Hermes Agent) can query this graph directly via Model Context Protocol (MCP), decreasing token spend by ~35%, reducing tool overhead/file lookups by ~70%, and avoiding deep file audits.
Uses web-tree-sitter and wasm grammars to index and parse symbols. Stores results in a local SQLite file database inside .codegraph/. It features a resident file watcher utilizing native OS event systems (FSEvents on macOS, inotify on Linux) to auto-sync modifications within a customizable 2-second debounce window.
Exposes tools that provide pre-indexed code understanding to the LLM. Instead of spawning broad "Explore sub-agents" that repeatedly call grep, glob, or read_file, the agent queries symbol definitions, impact radiuses, and call stacks directly via the local server.
A notable feature of CodeGraph is its custom resolvers that bridge boundaries static parsers typically miss:
Install the pre-compiled binary via curl (bundled Node runtime included, zero-dependency):
# macOS or Linux
curl -fsSL https://raw.githubusercontent.com/colbymchenry/codegraph/main/install.sh | sh
# Already have Node? Just use npm directly
npm i -g @colbymchenry/codegraph
Usage in an active software project:
# Initialize database and populate .codegraph/ index
codegraph init -i
# Force a manual sync check (though watcher auto-syncs by default)
codegraph status
CodeGraph features tree-sitter AST symbol resolution support covering 20+ technologies: