Caveman

1. What is Caveman?

Caveman is a set of rules that makes your AI coding assistant answer very briefly, in a "caveman" voice: no greeting, no warm-up sentences, just the point. AI services bill by the "token" (a unit of roughly three quarters of a word) for both what the AI reads and what it writes, so shorter answers cost less and are quicker to read.

To be plain about it: this is a tool for people who already use an AI coding assistant such as Claude Code, Codex, Gemini CLI, Cursor, Windsurf, Cline or Copilot. It is not a standalone app and it does not work with ChatGPT in a browser.

Caveman comes in two parts:

• The skill (small): one rule file that makes the AI talk less. Free, MIT licensed, no account or API key, and it works with more than 30 assistants.
• The proxy (large, optional): a program that runs on your machine between the assistant and the AI provider and shrinks what the AI has to read (logs, test output, JSON, diffs). Originals are always kept locally so the AI can pull them back. The proxy's engine is under the BSL-1.1 license: the source is readable and you may self-host it free for your own use.

Caveman does not shorten code, commands, file paths or error messages. Security warnings and confirmation questions are still written out in full.

A good fit if:

• You read your AI's answers a lot and find them wordy.
• You pay per token or keep hitting usage limits.
• Your sessions are long and heavy on logs and test output (that is when the proxy is worth it).

Not a fit if:

• You are billed per request rather than per token (GitHub Copilot premium requests, for example): a shorter answer is still one request.
• Your work is almost entirely code generation with little prose to cut.
• You only ask very short questions. The rule set itself adds about 1,000 tokens to every call, which can cost more than it saves.

2. What you need

• An AI coding assistant already installed. Caveman wraps the assistant you use; it does not replace it.
• Node.js: the runtime for JavaScript tools; it gives you the npx and npm commands. Caveman's full installer requires Node.js 22.13 or newer. Free download at https://nodejs.org
• Windows: PowerShell 5.1 or newer (included with Windows 10 and 11) if you use the full installer.
• No account, no API key and no payment for the skill.

The latest version at the time of writing is v2.7.0, released on 15 September 2026.

3. Step-by-step installation

Caveman has no .exe installer and no app store listing; every install route uses the command line. Start with the small part (Steps 1 to 3). The later steps are optional.

Step 1: Check Node.js

node --version

What success looks like: a version number starting with v. If the command is not found, install Node.js from https://nodejs.org and open a fresh terminal window.

Step 2: Install the skill with one command

npx skills add JuliusBrussee/caveman -g

The -g installs into your user-level skills folder, so it applies to every project. Without -g the skill belongs only to the current project. On the first run npx may ask permission to download the skills tool; say yes and follow the on-screen questions.

What success looks like: the installer reports that the skill was added for the assistant it found on your machine.

Step 3: Switch Caveman on in your assistant

Open your AI assistant. If it has not already started answering tersely, type:

/caveman

What success looks like: the assistant confirms in a very short sentence, and from then on its answers skip the preamble.

Step 4 (optional): The full installer

This finds every supported assistant on your machine, installs Caveman for each one in that assistant's native way, and for Claude Code also wires up "hooks" (small scripts that run when a session starts) and a status badge. It is safe to run more than once. It needs Node.js 22.13 or newer.

Windows (PowerShell):

irm https://raw.githubusercontent.com/JuliusBrussee/caveman/v2.7.0/install.ps1 | iex

macOS / Linux:

curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/v2.7.0/install.sh | bash

Note: Both commands download a script and run it immediately. To see what it would do first, on macOS/Linux you can run it in preview mode, which writes nothing:

curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/v2.7.0/install.sh | bash -s -- --dry-run

If you only use Claude Code, you can install for it alone with two commands:

claude plugin marketplace add JuliusBrussee/caveman
claude plugin install caveman@caveman

What success looks like in Claude Code: typing /caveman gets a clipped reply, and the status line at the bottom shows [CAVEMAN].

Step 5 (optional): Install the proxy so the AI reads less

Only do this once you are comfortable with the skill and have long working sessions. Run these two commands one after the other:

npm install -g @caveman-ai/cli
caveman setup --install

Then launch your assistant through Caveman. For Claude Code:

caveman claude

Other supported names: codex, gemini, aider, kilo, qwen, opencode, hermes, openclaw, pi.

What success looks like: the assistant opens as usual. From now on logs, test output, JSON and diffs are shrunk before they reach the AI provider. Your login with the provider passes through untouched.

4. First use

Ask the assistant any coding question, for example why a user-interface component keeps re-rendering. Before, you would get several sentences of explanation. With Caveman the reply is just the cause and the fix in a line or two, with function names and code left exactly as they are.

Commands worth trying inside the assistant right away:

Type	Result
/caveman lite	Tight, but still polite full sentences
/caveman ultra	Extremely short, close to keywords only
/caveman-commit	Writes a one-line commit message (the note saved with a code change)
/caveman-review	Code review with one finding per line
/caveman-compress CLAUDE.md	Shrinks the assistant's memory file, keeping every heading, path and command; the original is backed up
/caveman-help	A one-screen reminder of every mode and command
stop caveman	Turns it off; normal prose returns

If you installed the proxy (Step 5), the first thing worth doing is finding out where your tokens go:

caveman learn

It reads the assistant history already on your disk (read-only, locally) and ranks your biggest token sinks with a suggested fix for each. For history and usage estimates:

caveman stats

On savings, the author publishes the unflattering numbers along with the good ones. An independent JetBrains test on 86 real coding tasks, using the skill alone, found 8.5% fewer output tokens with no detectable change in quality. With the proxy, the repository's own benchmark shows 33.2% fewer input tokens overall, including one case that got 9.9% worse. Your results will differ; the author's advice is to measure on your own work.

5. Common problems and fixes

Symptom	Cause	Fix
Typing npx or npm gives "command not found"	Node.js is missing, or the terminal was not reopened	Install Node.js from https://nodejs.org and reopen the terminal
The full installer complains about Node.js	Node.js is older than 22.13	Upgrade Node.js and rerun. The installer does not install Node.js for you
Installed, but the assistant still writes long answers	The open session has not loaded the skill; in many assistants the skill is activated per session	Type /caveman. For Claude Code, restart it, because the hook only runs at session start
One window talks caveman and another does not	By design: the mode is stored per window	Type /caveman in the window where you want it
On Windows, the Step 4 install command is blocked by script execution policy	PowerShell restricts running scripts	Run Set-ExecutionPolicy -Scope Process -ExecutionPolicy Bypass in that same window, then rerun the install. The setting only applies to the open window
On Windows, hooks do not run	install.sh was used instead of install.ps1, or PowerShell is too old	Use install.ps1. Check the version with $PSVersionTable.PSVersion (5.1 or newer is required)
claude plugin update caveman says the plugin is not found	The plugin's full name includes its marketplace	Use claude plugin update caveman@caveman
Claude Code's settings.json is broken after installing	Something went wrong while merging settings	The installer writes a backup named settings.json.bak in the same folder before it changes anything; restore from that
caveman trial says it cannot measure	A trial needs its own proxy while the assistant is already wrapped	Run caveman disable claude, try again, then caveman enable claude afterwards

If the install breaks for no clear reason, the author suggests opening your own AI assistant inside the Caveman repository folder and asking it to read CLAUDE.md and INSTALL.md and install Caveman for you. If that fails too, report it at https://github.com/JuliusBrussee/caveman/issues

6. Uninstalling and updating

Updating

• Skill installed with npx skills add: rerun the exact command from Step 2; it overwrites in place.
• Claude Code plugin: claude plugin update caveman@caveman, then restart Claude Code.
• Anything the full installer set up: rerun the installer.

npx -y github:JuliusBrussee/caveman

Uninstalling

Order matters if you installed the proxy: run Caveman's uninstall first, and remove the CLI afterwards. The uninstall needs the CLI still present to restore each assistant's previous configuration.

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

Only then remove the CLI:

npm uninstall -g @caveman-ai/cli

That uninstall does not remove skills added with npx skills add. Remove the skill with:

npx skills remove caveman

Two leftovers have to be deleted by hand if you want a completely clean machine: rule files the installer copied into individual projects (only if you used the --with-init option), and the lifetime statistics file .caveman-history.jsonl in Claude Code's configuration folder.

7. Frequently asked questions

Does it cost anything? The skill is free (MIT). The CLI is MIT as well, while the proxy engine is BSL-1.1: free to self-host for your own use, with a commercial license required only if you host it as a service for third parties. You still pay your AI provider as usual.

Does it need an internet connection? Yes to install. The skill and hooks run entirely on your machine. Your AI assistant still needs the network to reach its model.

Is my data sent anywhere? Your assistant still talks to the provider you chose. The skill and hooks send nothing. The caveman CLI is different: the README says it sends anonymous usage statistics by default (which commands ran, plus token counts through and cut), never your prompts, code or file paths. Turn it off with:

caveman telemetry off

Or set the environment variable DO_NOT_TRACK=1. The exact boundaries are documented in the repository's SECURITY.md.

Does it make the AI less capable? Caveman only changes how the AI phrases things. Code, commands, paths and error messages are left intact. The JetBrains test found no measurable difference in quality.

Will it run on a weak computer? The skill is a text file and uses no resources of its own. The proxy is a background program on your machine; if you do not need it, skip Step 5.