What is Caveman?
Caveman started in April 2026 as a joke skill that tells a coding agent to drop filler and answer in clipped sentences while leaving code, commands and error messages untouched. Version 4 turned it into six parts installed by one CLI: the terse voice, an input compressor for logs, JSON, CSV and test output, an "Auto" model and effort router, reusable scripts called Blocks, a compressed web reader and a token-waste report.
It works with Claude Code, Codex, Gemini CLI, OpenCode and about 30 other agents, runs locally on your own keys or Claude Pro and Max login, and needs Node.js 22.13 or newer. Julius Brussee maintains it; a hosted Caveman Cloud for observability and evals sits alongside.
Is Caveman worth it?
The best independent test so far cuts the claim down to size. JetBrains ran the skill on 86 coding tasks with it forced on in every reply, about 240 billed trials in all, and measured 8.5% fewer output tokens (592,000 to 542,000 across 82 paired tasks) against the 65% the repository advertised. Quality did not move: a sign test gave p = 0.82. The reason is structural. Agent sessions are mostly code, diffs and tool calls, which Caveman deliberately leaves alone, so only the narration between them shrinks.
The maintainer's own numbers are more candid than the banner. Against a plain "Answer concisely." instruction, /caveman cut output by 3% at the median and /ultracave by 35%, on 10 questions in a single run, while the rules add about 1,160 input tokens. An honest-numbers page lists where it loses outright, including GitHub Copilot, which bills per request, so shorter answers save nothing.
The input compressor is the more useful half. On large CSV, log and YAML files it removed around 98% of tokens before the model read them, and across six file types a full Claude Code session used 33.2% fewer input tokens, though HTML came out 9.9% worse and that session run dates from August 2026 on an older engine. The Auto router claims 54% lower cost than Opus 5.5 at xhigh effort on 85 tasks; the README itself says the run is half done and the score gap is within noise.
The catch: the CLI sends usage statistics by default (commands, token counts, an install ID, OS and IP), and Auto sends your latest request with its attached files to Caveman Cloud, which may keep it on the free plan. With 119,975 stars against 2,042 commits, the star count tracks a viral moment more than engineering depth.
Our take: the voice is a style preference that costs nothing in quality; the compressor is where the money is. RTK covers the command-output side with fewer moving parts.
Where it earns its place
A contractor paying Claude Code API costs from their own pocket
Trim the bill on long debugging sessions full of test output and logs
Install the CLI and leave the input compressor on, so test runs and logs reach the model as short summaries. The repository measured 33.2% fewer input tokens per session on its test files; expect less on code-heavy work.
An engineering lead at a 30-person company evaluating token-saving add-ons
Decide whether a terse-output skill is worth rolling out
Run `caveman trial` on a real session, then compare against JetBrains' finding of 8.5% fewer output tokens with no quality loss. Turn telemetry off with `DO_NOT_TRACK=1` and leave Auto off if prompts must not leave your providers.
A developer whose agent reads large web pages
Stop browser snapshots from filling the context window
`caveman browse` returns a compressed page: the project measured 121 tokens instead of 15,704 for a 200-row table, though tiny forms come out 2.3 times larger.
Before you commit
- Setup
- `npx @caveman-ai/cli@latest` (Node.js 22.13+) finds installed agents and shows what it will change; the skill alone installs with `npx skills add JuliusBrussee/caveman -g`.
- Integrations
- Claude Code, Codex, Gemini CLI, OpenCode, Aider, Kilo, Qwen, Hermes, OpenClaw, Pi and more; middleware for the Vercel AI SDK, LangChain, OpenAI and Anthropic SDKs.
- Data and privacy
- The skill sends nothing. The CLI sends usage statistics by default (commands, token counts, install ID, OS, IP; not prompts or code); `caveman telemetry off` or `DO_NOT_TRACK=1` stops it. Auto sends your latest request and attachments to Caveman Cloud, which may keep it on the free plan.
- Cost at scale
- The tool is free; savings depend on billing. JetBrains measured about 8.5% fewer output tokens and roughly 10% lower cost at best on agent work, erased in one run by a single expensive task.
- Limits
- Rules add about 1,160 input tokens; HTML has no compressor yet; Blocks is a Python-only pre-release with no sandbox; per-request billing sees no saving.
What’s good
- No measurable quality loss in JetBrains' 86-task A/B
- Input compressor removes most noise from logs, CSV and JSON
- Code, commands and error text kept byte for byte
- Unusually frank page on where it loses
What’s not
- Real agent savings near 8.5%, not 65%
- Usage telemetry on by default
- Auto sends prompts and attached files to Caveman Cloud
- No saving on per-request billing such as Copilot
Best for
Developers paying per token for long agent sessions who want shorter replies and smaller tool output.
Caveman alternatives
Caveman: common questions
- Is Caveman worth it?
- Caveman is a harmless way to shave a little off an agent bill, as long as you believe JetBrains' 8.5% rather than the repository's 65%. We rate it 3.6 out of 5.
- How much does Caveman cost?
- Free and open source (Apache 2.0). Auto needs a free Caveman account; Caveman Cloud is a separate hosted service.
- What are the downsides of Caveman?
- Real agent savings near 8.5%, not 65%; usage telemetry on by default; auto sends prompts and attached files to Caveman Cloud; no saving on per-request billing such as Copilot.
- Who is Caveman best for?
- Developers paying per token for long agent sessions who want shorter replies and smaller tool output.
- How much does Caveman cost for a team or in production?
- The tool is free; savings depend on billing. JetBrains measured about 8.5% fewer output tokens and roughly 10% lower cost at best on agent work, erased in one run by a single expensive task.
- Is Caveman safe for company data?
- The skill sends nothing. The CLI sends usage statistics by default (commands, token counts, install ID, OS, IP; not prompts or code); `caveman telemetry off` or `DO_NOT_TRACK=1` stops it. Auto sends your latest request and attachments to Caveman Cloud, which may keep it on the free plan.
- What are the best Caveman alternatives?
- The closest alternatives we have reviewed are rtk (Rust Token Killer), OmniRoute and Claude Code Router.
Answers checked Nov 14, 2026.