Caveman

Caveman is a set of tools for making AI coding agents use fewer tokens. Its best known part is an open source skill: a single rule file that tells the agent to answer in compressed prose, dropping preamble while copying code, commands, file paths and exact error strings through unchanged. Alongside it sit a local proxy and compression engine that shrink what the agent reads, for example folding dozens of identical log lines into one marker while keeping the original retrievable on request. It is built for developers running Claude Code, Codex, Gemini CLI, Cursor, JetBrains Junie and similar agents, and for teams building their own agents, who can use middleware packages for TypeScript and Python with LangChain to compress tool results. The skill offers lite, full and ultra levels plus classical Chinese variants, can be switched off in session, and in Claude Code installs as a plugin with hooks that record the active mode. The documentation cites a JetBrains paired test of 86 coding tasks that found 8.5 percent fewer output tokens with no detectable quality change. The skill, the local CLI and the compression tools are free and work without a Caveman account; the skill is MIT licensed. Caveman Platform, a hosted gateway with model routing, caching, traces, evaluations, budgets and automated fixes delivered as pull requests, is in private development, and no prices are published. Access is by waitlist or a booked call. The documentation is candid about limits: the skill is sent as input tokens on every call, about 1,650 tokens for the base file, so on very short questions it can cost more than it saves, and per request billing gains nothing.