Caveman 🧩 Skill Free

What it does

Caveman cuts the tokens a coding agent spends, and it comes in two parts. The skill makes the agent answer in clipped, "caveman" prose. Code, commands, file paths and error messages stay exact, and security warnings and confirmations come back in full sentences. The optional proxy runs locally between your agent and the model provider. It compresses what the agent reads, such as logs, test output, JSON, diffs and search results. It keeps the originals in a local SQLite store so the agent can fetch the full version back.

What is inside

• The skill: intensity levels (lite, full, ultra and classical-Chinese variants), plus /caveman-commit, /caveman-review, /caveman-compress for shrinking CLAUDE.md-style files, and a few "work pattern" skills that encourage smaller changes.
• Subagent presets: compact investigator, builder and reviewer agents.
• The CLI and proxy: caveman claude (or codex, gemini, aider, opencode and more) wraps an agent. caveman learn ranks your token sinks from local history, caveman shrink compresses command output, and caveman trial runs A/B comparisons. There is also an MCP toolkit.

The README is candid about the numbers. An outside test found the skill alone cut output tokens by single digits with no measurable quality change, while the rules themselves add input tokens on every call.

Works with

The skill installs into 30+ agents, including Claude Code, Codex, Gemini CLI, Cursor, Windsurf, Cline, GitHub Copilot and OpenCode. The proxy has wrap profiles for about ten agents.

How to install

npx skills add JuliusBrussee/caveman -g

Maintenance and safety

Very actively maintained. The licence is split: the skill and CLI are MIT, while the proxy engine is BSL-1.1. You can self-host it for your own traffic, but hosting it for others needs a commercial licence. The CLI sends anonymous usage stats by default, which you can turn off with caveman telemetry off. The full installer is a curl | bash script that adds Claude Code hooks, so read it first. A hosted Caveman Cloud is on a waitlist.

Who should use it

Developers who pay per token and run long agent sessions full of logs and test output. It helps less if you are billed per request, as with Copilot premium requests.