About Caveman

One engine, five layers. The Caveman Skill is MIT and runs inside Claude Code, Codex, Cursor and 30+ other agents. The Caveman Proxy wraps an agent you already run and compresses eligible context on your machine, keeping the original bytes recoverable. The Agent SDK gives production agents a local price guard and a token bill. And Caveman Cloud is the managed plane where caching, compression and routing run behind eval gates. Cloud is in private development; the waitlist is open.

The open-source skill reached #1 on Hacker News and GitHub Trending, and now carries more than 74,000 stars. It is one prompt, doing one thing: strip the words a model does not need, and never touch the bytes that have to stay exact — code, commands, errors. Everything commercial here is that same idea moved closer to your traffic.

Spend is provider-reported usage priced against the public model catalog. It is a subtotal, not an invoice. Unknown models stay visibly unpriced rather than guessed or averaged. Savings climb a ladder — inferred, replayed, verified — and a number never skips a rung. Verified savings start at $0.00 and move only on provider-causal evidence. No fee, credit or discount is ever calculated from a savings figure, which is why we can afford to publish the honest zero.

Prompts and responses travel from your machine straight to your provider. Caveman's cloud is not in that path: it receives token counts, model names and savings numbers, never prompt or response bytes. On Enterprise deployments it receives nothing at all — zero data retention is enforced at write time and the uplink refuses the upload. The full table is on the data use page.

Caveman is built by Julius Brussee and a small team, working in public. The skill, the browser extension and the benchmarks are open source on GitHub; the research write-ups are in Labs and News.

Contact has every route in. The fastest one is contact@caveman.so.