Caveman cuts the tokens a coding agent spends in two places. A skill shortens what the agent says, and a local proxy shrinks what the agent reads. The two install separately, so you can take the output style alone or add the proxy on top.
The skill gives replies a fixed structure: answer first, one idea per sentence (modeled on the ASD-STE100 controlled English used in aircraft maintenance manuals), no greeting, plan announcement, or recap. Negations, numbers, units, code, commands, paths, and error messages stay verbatim, and security warnings, irreversible actions, and step-by-step instructions switch back to full sentences. Your own prompts are never rewritten. Three levels trade readability for length: /caveman, /ultracave, and /megacave, which answers in classical Chinese.
The proxy runs on your machine with your own keys or Claude subscription login and compresses logs, CSV, YAML, JSON, test output, and web pages before the agent reads them, keeping each original on disk for the agent to pull back. A CLI ranks where a session's tokens go from existing agent history, applies fixes on approval, A/B tests a real session, and compresses command output. Middleware packages for TypeScript and Python wrap Vercel AI SDK, LangChain, OpenAI, or Anthropic calls in your own agent.
Published measurements for the proxy, across three Claude Code runs per file type, with 18 of 18 answers correct:
| File type | File tokens after compression | Whole-session saving |
|---|---|---|
| CSV | 28,041 → 314 | 55.1% |
| Logs | 22,810 → 348 | 50.2% |
| YAML | 20,447 → 178 | 46.2% |
| Test output | 18,806 → 203 | 27.8% |
| JSON | 18,837 → 281 | 26.4% |
| HTML | no compressor yet | 9.9% worse |
| All six | 130,611 → 22,994 | 33.2% |
On ten developer questions, /caveman cut output tokens 3% below a plain "answer concisely" instruction at the median and /ultracave 35%. Outside evaluations include an Adobe Research paper reporting a 1.4 to 2.4 times cost cut across eight models and a JetBrains A/B test on 86 coding tasks reporting no measurable quality loss and 8.5% fewer output tokens. Adobe's finding that compressing the user's prompt hurts answers is why Caveman leaves prompts alone. The skill gains nothing where billing is per request rather than per token, and the CLI sends usage statistics by default, with an opt-out.
Features
- Output modes: /caveman, /ultracave, and /megacave shorten replies at three levels, and /caveman off restores normal output
- Verbatim payload: code, commands, paths, error messages, numbers, and negations are never dropped or altered
- Input proxy: compresses logs, CSV, YAML, JSON, test output, diffs, and web pages before the agent reads them, with originals kept on disk
- Token audit: ranks where tokens go from existing agent history and applies fixes one diff at a time on approval
- Compressed browsing: returns a compact version of a web page in place of a full Playwright snapshot
- Memory-file compression: /caveman-compress shrinks files such as CLAUDE.md while keeping headings, code blocks, and paths, and backs up the original
- Terse commits and reviews: one-line Conventional Commits and one-finding-per-line code reviews
- cavecrew subagents: subagents that find, edit, and review code and report back in the compressed style
- SDK middleware: TypeScript and Python wrappers for Vercel AI SDK, LangChain, OpenAI, and Anthropic calls
- Agent coverage: the skill works in Claude Code, Codex, Gemini CLI, Cursor, Windsurf, Cline, Copilot, and 30+ more agents
