Most LLM-based code explorers spend their budget repeatedly scanning files; that discovery phase dominates latency and token use. CodeGraph flips that model by building a local SQLite-backed semantic graph of symbols, calls, and routes so exploration agents can query relationships instantly instead of re-reading the repo.
What Sets It Apart
- Pre-indexed semantic graph: extracts symbols, call graphs, imports and framework routes so agents use a single graph query instead of many filesystem reads—this directly reduces tool calls and tokens used by Claude Code. (So what? Faster answers and lower token/tool-call costs for interactive exploration.)
- 100% local with FTS5: stores data in a local SQLite DB with full-text search and no external services or API keys. (So what? Suitable for sensitive codebases and offline workflows.)
- Always-fresh auto-sync: native file-watcher (FSEvents/inotify/ReadDirectoryChangesW) keeps the index current with debounced updates. (So what? Low maintenance—index stays current as you edit without manual re-indexing.)
- Cross-language & framework-aware: recognizes 19+ languages and common web-framework routes to connect URL patterns to handlers across languages. (So what? Enables multi-language call traces and route-aware impact analysis.)
Who it's for + Tradeoffs
Great fit if you use Claude Code (or an MCP-capable agent) to explore and change real codebases and need low-latency, repeatable lookups without sending data off-host. It's especially valuable for large repos where repeated grep/read cycles are costly. Look elsewhere if you only need occasional ad-hoc file searches (the indexing step adds upfront work), or if you require a hosted SaaS index or integrations outside the supported MCP tooling model. Also note native performance depends on a native SQLite backend—environments that fall back to a WASM SQLite will be slower.
Where It Fits
Acts as an on-machine MCP server and CLI/Library: install with npm, run codegraph serve --mcp to expose tools (search, context, callers, impact, node, files, status) to Claude Code explore agents. It is positioned as an AI-development infra tool that optimizes LLM-driven code exploration rather than a general-purpose code search service.
