Index a TS/React repo (monorepo or standalone) into a code graph and persist it to .code-graph/graph.json

Read the code graph. Defaults to a compact summary on large repos (pass full:true for everything). Narrow with type/depth/lean/fields to keep output small.

Read a single node from the persisted code graph by id

Reverse lookup: which components render the given component id (incoming renders edges)

Reverse lookup: which symbols call the given function/component id (incoming calls edges)

All references into a node (incoming edges), optionally filtered by edge type

Everything that transitively depends on a node — the impact if it changes or breaks

Find dependency cycles in the graph (import/render/call cycles)

Dead-code candidates: exported files/components/functions that nothing imports, renders, or calls. Excludes entry points (index/main/App/config) unless includeEntryPoints is set. Heuristic — review, do not auto-delete.

Fuzzy-find nodes by name or path (e.g. "useAuth", "Header") — resolve a rough name to canonical node ids without grepping or guessing ids

One dense bundle for safely editing a node: its source + what it depends on + what depends on it + blast-radius size. Use instead of get_node + source + who_calls + blast_radius.

Compute local vector embeddings for the indexed nodes (enables semantic_search). Idempotent + incremental — only changed nodes are re-embedded. No-op if the local model is unavailable.

Find code by MEANING, not just name (e.g. "logic that decides trustee access"). Ranks nodes by embedding similarity; falls back to lexical search (with a hint) if embeddings aren't built.

Start the local 3D web explorer + chatbot for a repo and return its URL. Use when the user asks to SEE/visualize the code graph, open the UI, or explore visually. Starts a background HTTP server on 127.0.0.1 (default port 3002) and indexes the repo; returns the URL to open in a browser.