
Caveman
Makes your agent say less, to save tokens.
New to this? What open source means and how to read this page.
What it is
A skill, with an optional proxy, that makes a coding agent answer in short, terse replies and read less, to cut the tokens it uses while keeping code exact.
Good for
Cutting the cost and noise of long agent replies.
The catch
The savings quoted in the README are the project's own numbers, so measure on your own work. The terse style is not for everyone, and the proxy sits between your agent and the model, so check what it handles.
Install it with an AI assistant
Paste this into Claude, ChatGPT or another assistant that can work on your computer (Claude Code or Claude Desktop, for example). It is written from the project's official repository and tells the assistant to show you every command first. Setup time is our estimate: about 3 min.
I want to install and try Caveman on my own computer. Official repository: https://github.com/JuliusBrussee/caveman Official site: https://docs.caveman.so/caveman License: Apache-2.0 Before you do anything: 1. Read the README and install notes in the official repository above, and nothing else. Do not follow third-party tutorials or mirrors. 2. Tell me in plain words what will be installed, what it needs (operating system, memory, graphics card, disk space, accounts) and what it will cost me, if anything. 3. Check my computer against those needs and tell me honestly if it will not work well. This is a skill or plugin for an AI coding agent. Before you install it, read its files and tell me which instructions, scripts and commands it contains and whether anything sends data out. Then install it for my agent (Claude Code) with the method in the README, and show me how to use it once. Rules: - Show me each command before you run it, and say what it does. - Ask before anything that deletes files, changes system settings, opens a network port or needs an administrator password. - Never put my passwords or keys on a command line, or in a file I did not choose. - When it is installed, help me check it with one small test, and tell me how to start and stop it. - If something goes wrong, explain it in plain words and do not guess your way around it. Good to know before you start: The savings quoted in the README are the project's own numbers, so measure on your own work. The terse style is not for everyone, and the proxy sits between your agent and the model, so check what it handles.
We have not run this prompt ourselves yet. Read what the assistant plans to do before you say yes.
Facts checked on 12 October 2026. We have not yet run a hands-on test of this tool. When we do, it gets a "Tested" mark here. Approximate US list prices for an individual plan, from third-party price trackers checked on 12 Oct 2026. Monthly billing where it exists. Regions, taxes and promotions change them. Confirm at the vendor before you decide.
More in Skills & repos
Anthropic's own public collection of Agent Skills.
A project worth knowing about
Beginnerabout 5 min
A software development method, packaged as skills.
A project worth knowing about
Beginnerabout 5 min
Anthropic's curated directory of Claude Code plugins.
A project worth knowing about
Beginnerabout 3 min