Caveman: a skill and proxy that cuts token usage for coding agents
Caveman is a skill and proxy for AI coding agents that trims token usage by making agents answer tersely and by shrinking logs, JSON and CSV before the model reads them. The author claims about 33% fewer input tokens and answers dropping from 63 to 20 tokens, plus cheaper Terminal-Bench 4 runs than Codex and Claude Code. It supports Claude Code, Codex, Gemini CLI and 30+ agents via a single npx install.
What this means for indie developers:If you use AI coding agents heavily, it is worth testing to cut token bills, but the savings and benchmark claims are self-reported and need verification.