If you want to reduce AI coding token usage, the most interesting experiment of 2026 is not a cheaper model — it is a proxy that talks like a caveman. Caveman is an open-source project that sits between your coding agent and the model, compressing what the agent reads and says. Its headline claim: up to 65% fewer tokens. This deep-dive looks at what the project actually does, what independent labs measured, how to install it, and whether your team should bother.
Who this is for: engineering leads and agencies running Claude Code, Codex, Gemini CLI or OpenCode at real scale, where API spend is a line item. By the end, you will know exactly which parts of Caveman save real money and which are theatre.
What Caveman actually is
Caveman (github.com/JuliusBrussee/caveman) started as a joke in April 2026: make your AI agent answer in caveman grammar so replies use fewer tokens. The joke went viral — #1 on Hacker News with 904 points, #1 on GitHub Trending, #8 on Product Hunt for the day — and then it got serious. As of today the repository has over 112,000 stars and 6,500 forks, written in Go and released under Apache-2.0.
Since version 4.0, Caveman is not one trick. It is a local proxy plus an agent skill with six switchable parts the author calls “rocks”: a compressed voice, input compression, a model router, reusable script blocks, waste analysis, and web-page compression. Install it with npx @caveman-ai/cli@latest (Node.js 22.13+), it finds your agents and logins, and your normal claude or codex command keeps working — now with Caveman in the middle. It runs on your machine with your own keys or Claude Pro/Max login, and it works with 30+ agents including Claude Code, Codex, Gemini CLI and OpenCode. If you only want the voice with no proxy, there is a skill-only install: npx skills add JuliusBrussee/caveman -g.
The six rocks that reduce AI coding token usage
1. Say less — the caveman voice
The voice is not broken grammar; it is a structured style built on ASD-STE100, the controlled English used in aircraft maintenance manuals: active voice, one idea per sentence, 20 words maximum. The same React answer that took 63 tokens takes 20 — “New object ref each render, so React re-renders. Wrap the prop in useMemo.” Crucially, the rules protect meaning: not, never, no and only are never dropped, numbers and units stay exact, and code, commands, paths and error messages pass through character for character.
The voice also knows when to stop: security warnings, irreversible actions and confused users get full sentences, then the grunt resumes. And it never rewrites your prompt — Adobe’s research found that cavemanning the user’s own prompt makes answers worse, so Caveman leaves your words alone. Measured on ten dev questions with claude-opus-5-5, plain output used 6,983 tokens, /caveman used 4,119, and the aggressive /ultracave used 2,693.
2. Read less — 33.2% fewer input tokens
This is the important rock, and the one most coverage misses. The voice shrinks what the agent says; this shrinks what it reads. Logs, JSON, CSV, YAML and test output are compressed before the model ever sees them. Across 54 Claude Code runs the project measured 33.2% fewer input tokens per whole session (885,793 down to 591,673), with 18 out of 18 answers still correct. A 28,041-token CSV came through as 314 tokens — a 98.9% saving. Test output went from 18,806 to 203.
To their credit, the red rows stay red: HTML has no compressor yet, so those sessions were 9.9% worse. The project ships an HONEST-NUMBERS.md file documenting every case where Caveman loses — a level of transparency we wish more AI products would copy.
3. Auto — 54% cheaper, same score
Auto is a new model in your agent’s picker that routes each task to the right model and effort level: hard ask gets the big brain, easy ask gets a cheap one. On 85 real coding tasks, Opus 5.5 at maximum effort solved 44.3% at $7.83 per try with a 35.2-minute median time. Auto solved 43.7% at $3.59 per try in 15.9 minutes — 54% cheaper in under half the time. Honest caveat from the project: the 0.5-point score gap is too small to call at that sample size, and most of the saving comes from Auto lowering effort on easy asks. Only Auto needs an account, and it is free (caveman login).
4. Blocks — code mode, script written once
Agents constantly write the same throwaway scripts: grep a test summary, peek at JSON, retry a flaky command. Blocks keeps those scripts in your repository so the agent reuses them instead of re-typing them. A go test -v ./... run whose raw output was 10,561 tokens came back as 76 tokens through the test-summary block, versus 397 tokens when the agent piped to tail -20 itself. Eight blocks ship with Caveman: json-peek, jsonl-stats, test-summary, first-error, wait-for, grep-defs, http-json and replace-in-file.
5 & 6. Waste fixes and Browse
caveman learn ranks where your tokens actually go using agent history already on your disk, and can apply fixes one diff at a time with your approval. A new warm-cache feature keeps Anthropic’s prompt cache alive while you step away, only when your own history says you come back in time. And caveman browse returns compressed web pages to the agent — 129.8× smaller, so a Playwright snapshot of a 200-row table costs 121 tokens instead of 15,704.
Why it is trending: the economics, not the meme
The meme got it on Hacker News; the economics kept it there. Agentic coding moved the bill from “a few dollars of chat” to “hundreds of dollars of autonomous sessions”, and three independent labs have now measured the core idea. Adobe Research published CAVEWOMAN on arXiv (June 2026): caveman-style output cut costs 1.4 to 2.4× per model, up to 3×, across eight models and five datasets. JetBrains ran 86 real coding tasks as a paired A/B in July 2026 and found no measurable quality loss (p = 0.82) alongside 8.5% fewer output tokens. Elastic rebuilt the idea for Elasticsearch and measured 63.6% fewer response tokens with, in their words, “zero information loss”. Even ThePrimeagen weighed in: “No way this actually works.”
The timing matters for business. If you are building AI integrations and automations for clients, or running agents across a dev team, token cost is now a genuine margin question. A tool that attacks the problem on both sides — smaller outputs and smaller inputs — is worth evaluating even if you never adopt the voice itself.
From our experience: measure cost per task, not tokens
At MTD Technologies we have spent nine years shipping software across hundreds of projects, and the pattern in the Caveman data matches what we see in real agent workflows: what the agent reads dwarfs what it says. Teams obsess over short prompts while their agents cheerfully ingest 28,000-token CSV dumps and 18,000-token test logs. That is why the “Read less” rock is the real product here and the voice is the marketing.
Two practical warnings. First, the project’s own honest numbers note the rules cost about 1,160 input tokens — if you pay per request (GitHub Copilot plans), shorter answers save you nothing. Second, do not chase the token number itself. The metric that matters is cost per completed task: a 65% token cut that breaks one build in twenty is a loss. Run caveman trial -- claude on your own work — the project ships an A/B command for exactly this — and keep the red rows in your own HONEST-NUMBERS equivalent.
Getting started in five steps
- Install:
npx @caveman-ai/cli@latest— the installer finds your agents, shows what it will change, and you press Enter. - Check status:
caveman statusshows which rocks are on and which agents they reach. - Find your waste:
caveman learnranks where your tokens go from history already on your disk. - Try Auto:
caveman login(free account) adds the Auto model to your pickers in Claude Code, OpenCode and Codex. - A/B it honestly:
caveman trial -- clauderuns a real session on your own work with and without Caveman.
If the evaluation looks promising but wiring it into your team’s workflow feels like a project on its own, that is exactly the kind of workflow automation work agencies handle every week. And if you want a wider view of the landscape, our AI Tools Directory tracks the tools we actually vet.
Frequently asked questions
Does Caveman really cut token usage by 65%?
The 65% figure is the project’s headline. The measured pieces are more specific: 33.2% fewer input tokens per session, a 1.4–2.4× cost cut in Adobe’s independent study, 54% cheaper task runs via the Auto router, and up to 98.9% compression on individual files like CSVs and logs. Your mileage will vary — the project says so itself in its benchmark docs.
Will the caveman voice make my agent dumber?
Independent tests say no, within the measured bounds. JetBrains found no measurable quality loss across 86 real coding tasks, and Caveman’s own runs got 18 of 18 answers right. The voice rules protect negations, numbers and code, and it never touches your prompt — only the agent’s output. Security warnings and irreversible actions still get full sentences.
Which agents does it work with?
Over 30 agents, including Claude Code, Codex, Gemini CLI, OpenCode, aider, Kilo Code and Qwen Code. The proxy runs locally with your own API keys or Claude Pro/Max login, so your traffic does not go to a third-party server. If you are not logged in to Auto, nothing is sent anywhere.
Is Caveman free?
Yes — Apache-2.0 licensed, free to install and use. The Auto model router needs a free Caveman account. The savings come from your own model spend: fewer tokens through the APIs you already pay for. Note that on per-request pricing (like some GitHub Copilot plans), shorter answers save nothing.
Who should not use it?
Teams that need verbose, audit-friendly agent transcripts for compliance; anyone on per-request pricing; and workflows where the agent mostly reads HTML (no compressor exists for it yet — those sessions measured 9.9% worse). Evaluate with caveman trial on your own tasks before rolling it out.
Caveman is a rare viral project that survived contact with independent measurement. Even if the voice is not for your team, its underlying insight is worth stealing: the biggest token waste in agentic coding is what the agent reads, not what it writes. Start there, and the savings follow. If you want help evaluating or deploying agent tooling across your team, talk to us — this is what we do.