One call
instead of
a crawl.
When an agent needs to understand code it greps, globs and reads one file at a time, rebuilding call paths by hand. CodeGraph hands it the relevant source, the call paths between those symbols, and the blast radius of a change in one call — from a local SQLite graph that re-syncs on every file change.
Install, wire your agents, index a project
codegraph install
cd your-project && codegraph init
Both are true at once
project’s own figuresThroughput
what it saves, per answer
Footprint
what it costs, per session
The same dense payload that makes it fast is what stays in the window. The project prints this column itself — that is why it gets the accent here.
Tool calls to answer one architecture question
with vs without · median of 4 runs per armBlack fill is the with-graph share of the without-graph call count. Figures are the project’s published re-measurement of 2026-08-05 on Claude Opus 4.8, on a harness that blocks the CLI in both arms; reproduced as stated, not re-run here. Cost savings vary far more than call counts — the project reports 13% on Django and roughly even on Gin.
Extract
A native Rust kernel parses source with tree-sitter grammars — functions, classes, calls, imports.
Store
Everything lands in a local SQLite database with FTS5 full-text search.
Resolve
Calls to definitions, imports to files, inheritance and framework patterns.
Auto-sync
OS file events, a two-second quiet window, incremental updates. Nothing to re-run.