# Caveman

> An open-source skill/plugin for Claude Code and 30+ AI coding agents that cuts output tokens by 65% by making agents respond in compressed, caveman-style prose while keeping code byte-exact.

Caveman is an open-source skill and plugin suite created by Julius Brussee that installs into Claude Code, Codex, Gemini, Cursor, Windsurf, Cline, Copilot, and 30+ other AI coding agents. It instructs agents to drop filler language and respond in tight, compressed prose — keeping code, commands, and errors byte-for-byte exact — while cutting output tokens by an average of 65%. The project reached #1 on Hacker News and GitHub Trending, and the repository has accumulated over 85,000 GitHub stars.

## What It Is

Caveman is a prompt-engineering skill that changes how AI coding agents communicate, not what they know. Once installed, the agent drops verbose explanations and filler phrases, responding in short fragments that preserve all technical substance. The project's benchmark table (committed in the `benchmarks/` directory and reproducible) shows output token reductions ranging from 22% to 87% across 10 representative coding tasks, with an average of 65%. The README explicitly notes that input and reasoning tokens are untouched, and the skill itself adds roughly 1–1.5k input tokens per turn, so whole-session savings run smaller than the output number alone.

## The Ecosystem of Tools

Caveman is the flagship of a five-tool suite, all built around the same idea — "agent do more with less":

- **Caveman** — the core skill; shrinks what the agent *says*
- **Caveman Code** — a full terminal coding agent built caveman-first, claiming roughly 2× fewer tokens than Codex on identical tasks, supporting 20+ providers with plan mode and an autopilot goal loop
- **Cavemem** — persistent memory served over MCP using local SQLite with FTS5 and a vector index, so agents stop re-sending context they already know
- **Cavekit** — a spec-driven build loop for compressed, structured development
- **CaveGemma** — the compression approach baked directly into Gemma model weights via fine-tuning

A browser extension for ChatGPT, Claude, and Gemini is also available on the Chrome Web Store for non-developer users.

## How It Works

Installation is a single shell command that auto-detects every supported agent on the machine and installs the appropriate skill file, plugin, or rule file for each. On Claude Code specifically, a hook writes a flag file each session so the agent talks caveman from message one without requiring a `/caveman` command. Six compression levels are available — `lite`, `full` (default), `ultra`, and `wenyan` (classical Chinese for maximum token density) — switchable mid-session with `/caveman <level>`. Companion commands include `/caveman-commit` for conventional commit messages, `/caveman-review` for one-line PR comments, `/caveman-stats` for real session token usage and lifetime savings, and `/caveman-compress` for rewriting memory files like `CLAUDE.md` into compressed form, which the README's receipts show cuts input tokens by an average of 46% on every subsequent session.

## Privacy and Architecture

The README states that Caveman has no telemetry, no analytics, no accounts, and no backend. After installation, zero network calls are made — the skill is a prompt, the hooks are local scripts, and `/caveman-stats` reads a log already on disk. The only network activity is at install time (fetching from GitHub and agent registries), documented in `SECURITY.md`.

## Update: v1.9.1 — "65%, honestly"

The latest release, v1.9.1, was published on July 3, 2026, under the tag "65%, honestly" — a nod to the project's commitment to transparent benchmarking. The repository was created in April 2026 and has been actively maintained since, with 4,778 forks and 367 open issues as of early July 2026. A managed cloud product (Caveman Proxy / Caveman Cloud) is in private development with a waitlist, described as a byte-safe LLM gateway that meters token spend to the cent without altering what the model sees.

## Features
- 65% average output token reduction across 30+ AI coding agents
- Six compression levels: lite, full (default), ultra, wenyan, and more
- Byte-exact preservation of code, commands, and errors
- Single install command auto-detects all supported agents
- Works with Claude Code, Codex, Gemini, Cursor, Windsurf, Cline, Copilot, and 30+ others
- /caveman-compress rewrites memory files for ~46% input token savings every session
- /caveman-stats shows real session token usage and lifetime savings in USD
- /caveman-commit generates conventional commit messages with ≤50-char subjects
- /caveman-review produces one-line PR comments
- caveman-shrink MCP middleware compresses tool descriptions
- Browser extension for ChatGPT, Claude, and Gemini
- No telemetry, no analytics, no backend — fully local after install
- Cavemem persistent memory over MCP with SQLite FTS5 and vector index
- Caveman Code full terminal coding agent with 20+ provider support
- CaveGemma compression baked into Gemma model weights

## Integrations
Claude Code, Codex, Gemini CLI, Cursor, Windsurf, Cline, GitHub Copilot, ChatGPT, OpenAI API, MCP (Model Context Protocol), OpenClaw, skills.sh registry, SQLite, npm

## Platforms
WINDOWS, MACOS, LINUX, WEB, API, BROWSER_EXTENSION, DEVELOPER_SDK, CLI

## Pricing
Open Source

## Version
v1.9.1

## Links
- Website: https://caveman.so/
- Documentation: https://caveman.so/docs
- Repository: https://github.com/JuliusBrussee/caveman
- EveryDev.ai: https://www.everydev.ai/tools/caveman