/caveman off or stop caveman. caveman disable for a wrap.

Caveman: pricing, features, alternatives

Caveman is a skill plus a local proxy that makes coding agents write less filler and read less noise. The repo JuliusBrussee/caveman was created 2026-04-04. On 2026-09-18 GitHub listed 106,284 stars and 6,155 forks. Docs live at docs.caveman.so. 106,284 stars is a heat signal, not traffic. GitHub SPDX is Other: the README splits MIT (skill, CLI surfaces) and BSL-1.1 (engine, proxy, related runtime), converting to Apache-2.0 on a stated change date. Treat that split as first-party, not a single OSI license.

The README's marketing line is "cuts 65% of tokens." Do not take that as a bill. JetBrains ran 86 coding tasks on the skill only (July 2026, before the proxy): 8.5% fewer output tokens, about 10% cost, no detectable quality change. Adobe Research's CAVEWOMAN paper measured caveman-style output cutting realized cost 1.4 to 2.4x across models, up to 3x in the best case. The repo's pinned 54-run Claude Code proxy suite reports 33.2% fewer provider input tokens overall, with one HTML case +9.9%. Measure on your own workload.

It pairs with Ponytail (less code) rather than replacing Claude Code. Cachebeat keeps a prompt cache warm; Caveman shrinks what goes through the cache.

Key Features

Skill (small rock): npx skills add JuliusBrussee/caveman -g. /caveman lite|full|ultra plus wenyan modes. Code, paths, errors, and security warnings stay full sentences.
Proxy (big rock): npm install -g @caveman-ai/cli then caveman claude (or codex, gemini, aider, opencode, pi, and others). Shrinks logs, JSON, diffs, and test output on your machine. Originals sit in local SQLite with a recovery handle.
Learn: caveman learn ranks token sinks from history already on disk. caveman learn implement proposes diffs one at a time.
Extra verbs: caveman shrink -- pnpm test, caveman browse <url>, caveman convert (pixel mode for skills), caveman stats.
Harness list: Claude Code, Codex, Gemini CLI, Aider, Kilo, Qwen Code, OpenCode, Hermes, OpenClaw, Pi. Wrap does not rewrite your config files.

Limitation: billed-per-request hosts (the README names GitHub Copilot premium requests) do not get cheaper from shorter answers. The skill adds input tokens on every call (README: about 1,000 estimated for the full skill). Native wrap for Codex skips the shrink hook because Codex rejects the rewrite. CLI telemetry is on by default; caveman telemetry off or DO_NOT_TRACK=1 stops it. Prompts and code are not in that payload per SECURITY.md.

Use Cases

Long sessions full of test logs and diffs, where reading dominates the bill.
People who like terse answers and still want the same diagnosis.
Skip if the job is almost all code generation with no prose, or you pay per request.

Pricing

Piece	Price	Notes from first-party pages 2026-09-18
Skill	$0	MIT. `npx skills add`. Works in 30+ agents per INSTALL.md.
CLI / proxy	$0 to self-host first-party traffic	MIT CLI. Engine/proxy BSL-1.1: self-host for your own traffic free; hosting for third parties needs a commercial license. About 106,284 stars.
Caveman Cloud	Waitlist	caveman.so. Signed receipts are a Cloud claim, not the local CLI.
Model usage	Your provider bill	Caveman does not sell tokens.

Getting Started

npx skills add JuliusBrussee/caveman -g and ask a coding question. Type /caveman if it does not wake.
Optional: npm install -g @caveman-ai/cli && caveman setup --install then caveman claude.
Run caveman learn before you tune anything.
caveman trial -- claude then caveman trial report if you need an A/B on your machine.

First-party start: quickstart and README.

Frequently Asked Questions

Does it make the model worse?

JetBrains reported no detectable quality change on the skill. The proxy suite checked 18/18 answers in that pinned run. Re-run on your tasks.

Will it rewrite my prompts into caveman-speak?

The README says no. Adobe found compressing the human prompt made answers longer and worse. Only the agent's mouth (and, with the proxy, what it reads) is compressed.

Can I turn it off?

/caveman off or stop caveman. caveman disable <agent> for a wrap.

Alternatives

Ponytail: shrinks what the agent builds, not what it says. The READMEs suggest pairing them.
Claude Code: the usual host. Caveman rides along.
Cachebeat: cache keepalive, not compression.

Tips

Start with the skill. Add the proxy only if caveman learn shows reading, not writing, as the sink.
Do not hide red benchmark rows from yourself. The HTML case in the wrap suite costs extra.
Turn telemetry off if that is a policy requirement.

Conclusion

Caveman is a joke that grew a skill, a proxy, and a research citation. Use the skill if you want shorter talk. Use the proxy if logs dominate the bill. Measure; the 65% headline is not your invoice.