Caveman
A free coding-agent skill and companion local proxy that rewrite an AI agent's prose into short, blunt phrasing to reduce the tokens it writes and reads, while leaving code, commands and safety warnings untouched.
Caveman is a skill for AI coding agents that rewrites the agent's prose into short, blunt sentences instead of long explanations, to reduce the number of tokens it writes. Code blocks, commands, file paths and exact error messages are left unchanged, and safety warnings or confirmation prompts are always returned in full sentences.
A companion local proxy runs between the agent and the AI provider and shrinks what the agent reads, such as logs, test output, JSON, diffs and search results, keeping a backup of the original content so it can be retrieved if needed. A middleware version applies the same approach inside custom code that calls LangChain, the Vercel AI SDK, OpenAI or Anthropic APIs directly, shrinking tool results before the model sees them while keeping the original in history.
The skill installs as a single rule file and works with more than 30 agent tools, including Claude Code, Codex, Gemini, Cursor, Windsurf, Cline and Copilot. It is MIT licensed.