codegraph

codegraph

Local code graph exposed as 8 MCP tools. Accelerates exploration with O(1) SQLite lookups and FTS5 search — slashing the grep + Read loops needed to orient in a codebase.

It does not replace grep. Codegraph is a tree-sitter approximation of the call graph, not a compiler-grade one: it is fast and low-token but unsound (it has false negatives). Treat it as the accelerator for orientation and traversal, and keep grep as the source of truth for completeness. See Soundness & limits below — this distinction is load-bearing, not a footnote.

When to use

• Brownfield repo where structure is non-obvious
• Task involves: "where is X used", "what breaks if I change Y", "trace flow from A to B"
• Long-running session — cold-start cost amortizes over multiple queries

When NOT to use

• Single-file edits or trivial lookups
• Session likely shorter than cold-start cost (see references/spike.md for per-repo numbers)
• Repo not initialized AND no cached .codegraph/ AND short session

Soundness & limits — read before trusting a result

Codegraph is built on tree-sitter, not the compiler, so its call/reference edges are an approximation. It fails in three confirmed ways. The governing rule falls out of them:

Grep is the source of truth for completeness. Codegraph is the accelerator for orientation and traversal. Never invert these.

The asymmetry is the whole argument: for a refactor, a false negative is dangerous (you miss a caller, ship a break); a false positive is merely annoying (you glance and discard). Codegraph has false negatives; grep does not. So codegraph may propose and rank, but grep confirms whenever the answer must be exhaustive.

The three failure modes (measured on a ~580-file Rust+TS repo):

1. Unsound edges → false negatives. callers/callees/impact miss real call sites when the receiver type can't be resolved from syntax alone (calls on locals, generics, trait objects). Measured: a method with 8 real callers returned 2 — the 5 dropped were local.method() and generic-dispatch sites, with no warning. Exact on unique free-functions + direct calls (verified 1/1, 2/2); unreliable on method dispatch.
2. Name resolution is fuzzy and overload-blind. Bare-name queries prefix-match (useWorkspace answered for useWorkspacesQuery — a different symbol) and conflate overloads (execute, with 19 definitions, returned sqlx's .execute() DB calls, not the trait method). For any name with multiple definitions, drop to codegraph_node on a specific symbol id; bare callers/callees are unattributable.
3. The index lags the disk. CLI sees nothing until codegraph sync (~0.3s); the MCP file-watcher closes the gap to ~1–2s. Right after an edit, the index is wrong — grep/Read is ground truth until it catches up.