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

I’m building Kivgraph, a local semantic code graph for AI coding agents. It maps symbols, dependencies, and cross-repository relationships so an agent can answer structural questions like “where is this used?” or “what breaks if I change it?” without repeatedly reading the same files.

In my benchmark, Kivgraph and grep + reading reached 28/29 exact answers, while the graph used 36k tokens versus 268k. Grep was still cheaper for some simple searches, so I see this as a complement rather than a replacement.

The launch is scheduled for tomorrow. I’d love to hear: what codebase questions make your agent burn the most context, and what would you want a tool like this to expose?