Caveman is a prompt-engineering skill that reduces AI coding agent output tokens by approximately 65% by forcing replies into condensed 'caveman-speak'. It preserves code and commands while dropping filler text, and ships with an MCP middleware tool for compressing MCP tool descriptions.
Project overview
The project addresses output token costs and readability directly at the prompt layer by compressing agent responses while maintaining byte-for-byte code and command accuracy.
Project type
MCP · Prompt Engineering
Use cases
Coding & Development
Deployment
Refer to project documentation
License
MIT
Best for
Developers using AI coding agents who want to reduce natural language verbosity and output token usage during CLI interactions.
Key capabilities
Installs a skill or plugin into AI coding agents that forces them to drop filler text and reply in condensed phrasing, saving output tokens while maintaining technical accuracy.
Provides six levels (lite, full, ultra, wenyan) that users can switch between anytime to control the exact compression style applied to the agent's output.
Ships an MCP middleware tool (caveman-shrink) that wraps any MCP server to compress its tool descriptions.
Limitations and risks
Caveman only shrinks output tokens; input and reasoning tokens remain untouched.
GPU and minimum hardware requirements are not documented.
Getting started
Setup is rated as easy. The specific first success path is not detailed in the available documentation.
Evidence and sources
GitHub project description: 🪨 why use many token when few token do trick
README: Caveman is a skill/plugin for [Claude Code](https://docs.anthropic.com/en/docs/claude-code), Codex, Gemini, Cursor, Windsurf, Cline, Copilot, and 30+ other agents.
README: Agent drops the filler and answers in tight caveman-speak, keeping code, commands, and errors byte-for-byte exact. You save output tokens on every reply, forever.
README: works_with-30%2B_agents-orange
README: ~30 seconds. Needs Node ≥18. Skips agents you no have. Safe to re-run.