The pitch for a GitHub repo with about 120,000 stars is written in grunts: “Why use many token when few token do trick.” The repo’s description claims it cuts 65% of tokens by making your AI agent talk like a caveman. JetBrains tested it on real coding work and measured 8.5%.
That gap is the whole story. Here is how to install the Caveman skill, what it does to your agent’s replies, and what it actually saves.
What you need
- A coding agent. The README lists Claude Code, Codex, Gemini CLI, OpenCode and “30+ more.”
- Node.js 22.13 or newer for the full installer, per the README.
- A willingness to read what you install. As of 15 November 2026 the repo (
JuliusBrussee/caveman, Apache-2.0) shows about 120,000 GitHub stars and was created on 4 April 2026.
These commands are copied from the project’s README as of 15 November 2026 and checked against it. We did not install the skill for this article, so the README stays the reference if anything changes.
Step 1: Choose how much to install
Caveman now comes in layers. The README describes six parts: shorter replies, shrinking logs and files before the model reads them, model routing, reusable scripts, waste fixes and compressed web pages. Only the first part is the famous caveman voice.
The README says the full installer finds your agents and logins and shows what it will change before you press Enter. It also says the CLI sends usage stats by default (commands run, token counts, a random install ID, OS and IP, never your prompts or code). If you only want the voice, take the smaller route.
Step 2: Run the install command
Skill only, no proxy:
npx skills add JuliusBrussee/caveman -g
Claude Code plugin:
claude plugin marketplace add JuliusBrussee/caveman && claude plugin install caveman@caveman
Full installer (finds your agents for you):
npx @caveman-ai/cli@latest
If you already have Claude Code open, the README gives the slash form: /plugin marketplace add JuliusBrussee/caveman, then /plugin install caveman@caveman. Our guide to Claude Code agent skills explains what a skill file is and why it deserves a skim before you trust it.
Step 3: Switch it on
Type /caveman ($caveman in Codex) if it does not start on its own. There are two stronger modes, /ultracave and /megacave (the README shows the latter answering in classical Chinese). Say stop caveman to go back.
Expected result, from the README’s own example. A normal answer (“The reason your React component is re-rendering is likely because you’re creating a new object reference on each render cycle…”) runs 63 tokens. The caveman version is: “New object ref each render, so React re-renders. Wrap the prop in useMemo.” That is 20 tokens. Code, commands and error messages stay untouched.
Step 4: Measure it on your own work
The README ships caveman trial -- claude to A/B a real session. The repo’s honest-numbers page gives a simpler rule: run the same task with and without Caveman, compare your provider’s billed totals, and if Caveman raises the cost, turn it off for that workload.
What it really saves
| Test | Result |
|---|---|
| Repo description | “cuts 65% of tokens” |
| JetBrains, 86 SkillsBench agent tasks | 8.5% fewer output tokens, no measurable quality loss (p = 0.82) |
| Repo’s own test, 10 dev questions on Opus 5.5 | 6,983 output tokens with no instruction, 4,334 with “Answer concisely.”, 4,119 with /caveman |
| LemonCrow, cited by the repo | Billed output down 44.1%, total cost up 0.06% |
Three things make the numbers smaller than the pitch.
- Agents mostly write code, not chat. JetBrains says code, diffs and tool calls dominate an agent’s output, and Caveman leaves them verbatim. Only the narration between tool calls shrinks.
- Politeness was already optional. The repo’s own test shows that a plain “Answer concisely.” gets most of the way. The README says, “‘Be concise’ is the real baseline; new models already know it.”
- Output is a small slice of the bill. One user’s test, quoted on the repo’s honest-numbers page, found output was 0.3% of all tokens across 662 Claude Code sessions.
JetBrains also warns that its figure is a ceiling. It forced the skill on in every reply, so normal use “can only be equal or lower.” Its first 10-task run showed a 30% saving that vanished at scale. JetBrains’ verdict: use it if you like it, “it costs you nothing measurable in quality,” but do not expect huge savings.
To be fair to the project, it publishes this itself. Its honest-numbers page opens with “Caveman save tokens sometimes. Caveman cost tokens sometimes,” and lists cases where it loses, such as agents billed per request (GitHub Copilot) where a shorter answer is still one request.
A separate paper, which the README credits to Adobe Research (arXiv, 23 June 2026, a preprint), found that caveman-style output cut cost 1.4 to 2.4 times per model for most API models across eight models. Compressing the user’s own prompt, though, raised net cost by roughly 1.15 times on average. That is why the README says Caveman never rewrites your prompts.
The bigger rocks
The newer parts of the repo target the real cost: what the agent reads. The README reports 33.2% fewer input tokens across 54 Claude Code runs on its test files, and a 54% cheaper “Auto” model router on 85 coding tasks, which it labels interim and half done. These are the project’s own measurements, and the README flags that HTML got worse. A “Browse” part also compresses web pages for the agent, which pairs with giving your agent a real browser through Chrome DevTools MCP.
Common errors and fixes
- Caveman does not start. Type
/cavemanby hand. - You had Caveman before version 4.0. The README says to run it with
@latest, or the old copy on your machine wins. - npm blocks the git install. On npm 12 or newer, the README gives
npx --allow-git=root -y github:JuliusBrussee/caveman#v4.1.0. - Costs went up. Turn it off with
/caveman off. To uninstall the skill:npx skills remove JuliusBrussee/caveman -g. - You do not want telemetry. The README says
caveman telemetry offorDO_NOT_TRACK=1disables it.
The verdict
Caveman is a cheap, fun style switch that, per JetBrains, does little harm to quality. If token costs are what worry you, start by choosing the right model; our Claude Fable 5.2 explainer is a good place to begin. Treat the 65% as a chat-answer number, and the high single digits as the agent-work number.
Sources
Links last checked on 15 November 2026.





Write a response