Caveman 4.0.0 reorganises the project into six parts, adding a local compression proxy and a model-and-effort router alongside the caveman voice
Caveman 4.0.0, released on 13 October, makes the caveman voice one of six parts in a repo that has passed 111,000 GitHub stars. The repo's Honest Numbers page, as dev.to reports it, puts that voice at about 3% over "Answer concisely".
The Engineer · Build desk
In the project's showcase example, an agent's explanation of a React re-render takes 20 tokens in caveman mode, down from 63 tokens before. The /ultracave setting drops it to 14 tokens, and /megacave answers in Classical Chinese for 13 tokens.
| Figure | Value | Note | Claim |
|---|---|---|---|
| React re-render explanation, caveman mode | 20 tokens | 11 | |
| Same answer, before caveman | 63 tokens | 11 | |
| /ultracave | 14 tokens | 14 | |
| /megacave, Classical Chinese | 13 tokens | 14 |
What happened
- The other parts are a local proxy that compresses what the agent reads, a per-request router for model and reasoning effort, reusable scripts, and tools for finding where tokens go.
- GitHub labels the repo Go because most of the code is now the compression engine and proxy, while the voice skill that made it famous is still a markdown file.
- After a full install all six parts are on, and only the Auto part needs an account, which is free.
- The voice has three intensities: the base setting answers the showcase example in 20 tokens, /ultracave in 14, and /megacave in Classical Chinese in 13.
Why it matters
- contradiction The famous 20-token example and the 3% figure use different baselines, so neither predicts the saving on a particular bill.
- decision Teams that want only the voice can install the markdown skill, which leaves the proxy and router out of the request path.
- exposure After a full install, plain claude, Codex, Gemini CLI and OpenCode all run through Caveman, so a fault in the compression path would reach every one of them.
The showcase example cuts an ordinary agent reply from 63 tokens to 20 [11], a drop of about 68% [16]. The Honest Numbers page, as dev.to describes it, compares the voice with a one-line instruction, "Answer concisely", and finds about 3% [4]. For the 68% to transfer, your agent's default replies would have to be as padded as the example, and replies would have to be a large share of your token spend. For the 3% to transfer, your workload would have to resemble whatever Honest Numbers measured. In our view that makes the voice a small setting once a concision instruction is already in place.
The page itself is good practice. According to dev.to, it measures the project's most famous claim and publishes the result even where it is unflattering [5]. The voice rules show the same care. Not, never, no and only are never dropped, because compressing "do not run this against production" carelessly gives "run against production" [12]. Code, commands, paths and error messages stay character for character [13].
Version 4.0.0 moves the effort to the input side [2]. The voice shapes how the model answers. The proxy compresses what the agent reads, and the router picks the model and reasoning effort for each request [3]. We'd expect those two to move more than phrasing does. The write-up does not give the workload behind the 3% or a measured saving for either the proxy or the router.
The proxy also needs the guarantee the voice already has. In our view, negations and verbatim payloads in what the agent reads should survive compression to the same standard, and its docs are where we would look first.
Defaults set the adoption cost. The installer finds the agents and logins on the machine, shows what it is about to change and waits for Enter. It needs Node.js 22.13 or newer [7]. Each part has its own switch, for example `caveman off browse`, and `caveman status` shows what is running [9].
What to watch
- A published measurement of the proxy's or router's token savings, with the workload and baseline named, alongside the voice figure on Honest Numbers.
- Documentation showing that the proxy preserves negations and verbatim code, commands and paths in what the agent reads, as the voice rules do for output.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence42
- Adoption70
- Hype gap+20
- Incentives
- Insufficient
- Confidence60
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
The Caveman repository has passed 111,000 stars and 6,400 forks.
- [2]
Version 4.0.0, released on 13 October 2026, reorganises the project into six parts the README calls "rocks", and only one of them is the caveman voice.
- [3]
The other parts are a local proxy that compresses what the agent reads, a router that picks the model and reasoning effort per request, a library of reusable scripts, and smaller tools for finding where tokens go.
- [4]
The repo's own Honest Numbers page says the caveman voice saves about 3% over "Answer concisely".
- [5]
The Honest Numbers page measures the project's most famous claim and publishes the result even where it is unflattering.
- [6]
GitHub labels the repo as Go because most of the code is now the compression engine and proxy behind the input-side parts; the CLI and installers use JavaScript and TypeScript, the evaluation harness uses Python, and the skill that made it famous is still a markdown file.
- [7]
The full install is the command npx @caveman-ai/cli@latest; it finds the agents and logins on the machine, shows what it is about to change, waits for Enter, and needs Node.js 22.13 or newer.
- [8]
All six parts are on after install, and only the Auto part needs an account, which is free.
- [9]
Each part has its own switch, for example caveman off browse, and caveman status shows what is running.
- [10]
The voice alone installs as a skill with no proxy and no CLI, using npx skills add JuliusBrussee/caveman -g.
- [11]
In the project's showcase example, an agent's explanation of a React re-render goes from 63 tokens to 20 in caveman mode.
- [12]
The voice rules never drop not, never, no and only, because compressing "do not run this against production" carelessly gives "run against production"; the rules protect those words explicitly.
- [13]
The voice rules leave code, commands, paths and error messages character for character.
- [14]
The voice has three intensities: /caveman gives the 20-token answer, /ultracave drops to 14 tokens, and /megacave answers in Classical Chinese for 13 tokens.
- [15]
After the full install, plain claude runs through Caveman, as do Codex, Gemini CLI, OpenCode and, per the install matrix, more than 30 other agents.
- [16]
The showcase example is a cut of about 68% in tokens (63 to 20).
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
1 independent publisher whose own reporting we read for this story.
- dev.toCaveman Claude Code Skill: What It Really Saves
1 article · October 15, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.
Topics
- Token Optimization TacticsFollow
- Open Source Developer ToolsFollow
- AI Coding AgentsFollow