Caveman

// Overview

Caveman attacks a specific, boring cost: AI coding agents bill by the token and write like they know it. It ships in two pieces. The "small rock" is a skill — a rule file that makes the agent answer in clipped, caveman-style prose. Code, commands, file paths and exact error messages are never rewritten; only the explanation around them is. The README's ten-prompt comparison against the Claude API puts average output tokens at 294 with the skill versus 1214 without, an average saving of about 65%, ranging from 22% on a task that was mostly code to 87% on one that would have produced an essay.

The "big rock" is a local proxy that sits between the agent and the model provider and shrinks what the agent *reads* before each call — usually the larger half of the bill. Everything the proxy squeezes is backed up on disk so the agent can pull the original back when it needs it. The two stack: most people install the skill first and add the proxy later.

Licensing is split. The skill and the CLI are MIT; the Engine-linked runtime directories (engine, proxy, rewriter, browse, mcp, shrink, the cavemem Go core and shared/platform) are Business Source License 1.1, and new Engine-linked runtime modules default to BSL unless explicitly classified as MIT. The skill installs into 30+ agents — Claude Code, Codex, Gemini CLI, Cursor, Windsurf, Cline, Copilot and more — and the proxy ships wrapper profiles for ten native agents.

// What it does

• A rule-file skill that compresses an agent's prose output while leaving code, commands, paths and error text untouched
• A local proxy that compresses what the agent reads before each provider call, with an on-disk backup of everything it squeezes
• Installs into 30+ agents via `npx skills add`, with a Claude Code plugin marketplace entry and a Gemini CLI extension
• Wrapper profiles for ten native agents (claude, codex, gemini, aider, kilo, qwen, opencode, hermes, openclaw, pi)
• A reproducible benchmark in the repo — `uv run python benchmarks/run.py` regenerates the ten-prompt token table
• Full installer wires Claude Code hooks and a statusline badge, is safe to re-run, and skips agents you do not have

// Getting started

Start with the skill — it is one command and free forever. Add the proxy when you want the input side compressed too. Node.js 22.13+ is required for the full installer.

npx skills add JuliusBrussee/caveman

npm install -g @caveman-ai/cli && caveman setup --install
caveman claude   # or codex · gemini · aider · kilo · qwen · opencode · hermes · openclaw · pi

claude plugin marketplace add JuliusBrussee/caveman && claude plugin install caveman@caveman
gemini extensions install https://github.com/JuliusBrussee/caveman
npx skills add JuliusBrussee/caveman --skill '*' -a codex --yes

uv run python benchmarks/run.py

npx -y github:JuliusBrussee/caveman -- --uninstall

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Cut the output-token bill on a coding agent you run all day, without changing the model or the quality of its diffs
• Compress large tool outputs and file reads on the input side before they reach the provider
• Standardise terse agent output across a team that uses several different CLI agents
• Measure what prose actually costs you by re-running the repo's ten-prompt benchmark on your own workload

// How Caveman compares

Caveman alongside other open-source agent skills & plugins tools AI/TLDR tracks, ranked by GitHub stars.