Caveman proxy claims 33.2% fewer billed input tokens

Caveman now cuts provider-billed input tokens 33.2% across more than 30 agents, per the project's GitHub page. That's real savings for anyone running agents against a metered API, where cost scales with everything fed into the model. The project has climbed to 98.9k GitHub stars and is trending.

Two pieces here. A skill compresses model output, averaging around 65% fewer tokens. One cited example: a React re-render explanation shrank from 1,180 tokens to 159.

A proxy handles the input side, compressing by content type. It applies structural compression to JSON, filters INFO and progress logs, and uses tree-sitter to elide function bodies in code. A pixel mode renders dense text to PNG for vision models to read instead.

Original bytes get written to a content-addressed recovery store before any lossy transform runs, so exact retrieval stays possible. That's not a detail you add casually. It suggests the team already hit cases where compressed output wasn't enough, or the original was needed back.

Install is one command: npm i -g @caveman-ai/cli, then caveman setup --install. The GitHub page doesn't say how the 33.2% figure was measured, on what workloads, or against what baseline. I'd want to see the compression fail on a real edge case before wiring it into a production agent setup. So far I've only seen the headline number.