268k tokens to understand 37 repos — Kivgraph did it in 36k

[Luqueee]

I built Kivgraph after watching AI coding agents repeatedly search through the same repository to answer questions like “where is this used?” and “what breaks if I change it?”

Kivgraph builds a local semantic graph of symbols, dependencies, and cross-repository relationships, so agents can navigate structure directly instead of reconstructing it from repeated file reads.

In a benchmark covering 37 repositories and 29 questions, Kivgraph and grep + reading both reached 28/29 exact answers. The graph used 36k tokens versus 268k for grep + reading — about 7.4× less context. Grep was still cheaper for simple searches, so this is a complement rather than a replacement.

It’s Apache-2.0, runs locally, needs no API key, and supports Go, TypeScript, Rust, Python, and Dart.

GitHub: https://github.com/Luqueee/kivgraph
Website: https://kivgraph.dev

I’d love feedback: what codebase questions make your coding agent burn the most context?