Caveman: A Token-Saving Skill and Proxy for Coding Agents
A skill and local proxy that make your coding agent say less and read less: same answers, about a third fewer tokens, free on the plans you already pay for.
Project intro
If you let a coding agent loose on your project, the bill is mostly words: long answers, and even longer logs, test output and files it has to read before it can do anything. Caveman is a free, open-source tool that shrinks both sides — the agent says less and reads less — and the code stays exactly the same.
It comes as two parts that work together. A skill teaches your agent to answer first and keep it short — one idea per sentence, no greetings, no recap — and a local proxy squeezes logs, JSON and test output down before the model ever sees them. In the project's own tests, whole sessions used 33.2% fewer input tokens, and a typical fix went from 63 tokens of answer to 20, with the same result.
It is not just the author's own numbers: Adobe Research cited it and measured costs cut 1.4 to 2.4 times, JetBrains ran 86 real coding tasks against it and found no measurable drop in quality, and it works with more than 30 agents, from Claude Code and Codex to Gemini CLI and OpenCode. One command installs it — npx @caveman-ai/cli — and it runs on your machine with the logins and subscriptions you already have.
There is also a newer free desktop app, Caveman Code, that puts every agent you use in one window. With nearly 120,000 stars on GitHub and fresh code landing on 14 November 2026, this is one of the most talked-about tools in AI coding right now — and it is free, so the only cost of trying it is a few minutes.
Actions
Tech stack
Project info
- Project category
- AI Agents
- Project status
- Online
- Collaboration intent
- None