CodeGraph

SQLite knowledge graph of a codebase's symbols, edges, and files (30+ languages), served over MCP. One codegraph_explore call returns the verbatim, line-numbered source of the relevant symbols plus the call paths between them and a blast-radius summary.

Setup (per project)

codegraph init     # builds .codegraph/ index in cwd (needed before MCP tools work there)
codegraph status   # index health; pending-sync files are listed

Auto-sync keeps the index fresh: the OS-native file watcher + 2s debounce, per-file staleness banners, and catch-up reconciliation on MCP reconnect. Manual codegraph sync is only needed with CODEGRAPH_NO_DAEMON=1 or CI pre-flight.

MCP tools

Intent	Tool
Almost any question — architecture, flow, bug, investigation	codegraph_explore ⭐ (primary; usually the only call needed)
Locate symbol by name	codegraph_search
Call chain	codegraph_callers / codegraph_callees
Change impact	codegraph_impact
Single symbol full source / overload disambiguation	codegraph_node
File structure	codegraph_files
Index health	codegraph_status

Principles

• codegraph_explore first — it groups source by file with line numbers and links the call paths.
• Answer directly from the graph; do NOT delegate exploration to file-reading subagents — that defeats the index.
• One explore beats search + node chained calls.
• Never use grep to verify codegraph results — AST edges are stronger evidence than text matches.
• Graph edges are proven, not complete: a missing edge does not prove a reference is absent.
• After edits, watch the staleness banner (⚠️ names a pending file) — Read that file directly for live content.