Your coding agent talks too much and reads too much, and you pay for both. Caveman is a free add-on that makes it answer in short, plain sentences, and shrinks logs, JSON and test output before the model reads them. It started as a joke in April 2026. It now has more than 113,000 GitHub stars.
Just want shorter answers? Read What It Is, then take the one-line skill install in Get It. Want the bigger savings? Read Six Parts and Before You Install first. The full install does a lot more, and some of it talks to a server.
Caveman trims what your agent says and reads, and you save on every session. Running a lean agent is a real skill. So is knowing when it's good enough to stop tuning and start selling.
I'm building a place where you learn to run your coding agent well on your own, and point the time you save at a product people pay for, with a CTO who's built two seven-figure software companies in your corner.
Tune Less. Ship More.
Free to join the waitlist. Opens at $49/mo, and the price goes up as seats fill, so first in pays the least.
Caveman began as a skill: a set of writing rules your agent follows. Answer first, one idea per sentence, no greeting, no recap. Code, commands, file paths and error messages are never touched. Version 4.0 adds a local proxy that sits between your agent and the model, so it can also shrink what the agent reads, pick a cheaper model for easy asks, and reuse scripts the agent already wrote. Each part has its own on/off switch.
The reason your React component is re-rendering is likely because you're creating a new object reference on each render cycle. When you pass an inline object as a prop, React's shallow comparison sees it as a different object every time, which triggers a re-render. I'd recommend using useMemo to memoize the object.
New object ref each render, so React re-renders. Wrap the prop in useMemo.
The repo calls them rocks. All six are on after the full install. Numbers are the maintainer's own, from the README.
Plus a few extras in your agent: /caveman-commit writes one-line commit messages, /caveman-review gives one finding per line, and /caveman-compress CLAUDE.md slims your memory file and keeps a backup.
You don't have to take all six parts. Start small and add the rest if you want it.
One command, no proxy, no account. The rules live in your agent and nothing leaves your machine. Easy to remove if you don't like it.
Get It · Step 1Same voice, but it starts on its own in every session, subagents too. You also get the slash commands like /caveman-commit.
That's where most of your tokens go, and the voice can't touch it. The full install adds Read Less, Blocks, Browse and Auto across every agent it finds.
Get It · Step 3GitHub Copilot bills premium requests. A shorter answer is still one request, so the voice saves you nothing there. The repo says so itself.
Check Your Plan FirstFree on GitHub at JuliusBrussee/caveman. Works on Mac, Linux and Windows, with Claude Code, Codex, Gemini CLI, OpenCode and 30+ more.
npx skills add JuliusBrussee/caveman -g
Type /caveman if it doesn't start on its own. Say stop caveman to go back.
/plugin marketplace add JuliusBrussee/caveman /plugin install caveman@caveman
On claude.ai or the desktop app, add it under Customize > Plugins. The repo hasn't fully tested that path yet.
npx @caveman-ai/cli@latest
It finds your agents, shows what it will change, and waits for Enter. Keep the @latest, or an old copy wins.
Try this first: after the full install, run caveman trial -- claude on a real task. It A/B tests the session on your own work, so you see your number, not theirs.
/caveman/ultracave is shorter still. /caveman off stops it.caveman statuscaveman off browseoutput, input, waste-fixes, routing and scripts.caveman learncaveman statscaveman doctorcaveman telemetry offIn the repo's own test, just saying Answer concisely. cut output from 6,983 tokens to 4,334. /caveman took off 3% more on top of that. /ultracave took off 35%. The rules themselves add about 1,160 input tokens.
One user measured 662 Claude Code sessions: output was 0.3% of all tokens. That's why Read Less matters more than the voice. My Token Diet guide makes the same point.
Commands, token counts, a random install ID, your OS and IP. Not your prompts, code or file paths. Turn it off for good with caveman telemetry off or DO_NOT_TRACK=1. The skill on its own sends nothing.
Only if you log in and pick Auto. Each ask sends your latest message plus anything attached to it, like CLAUDE.md or files, and the end of the last reply. On the free plan they may keep it. caveman off routing stops it.
The Auto benchmark is half done. The 33.2% session test ran on an older engine, and HTML files came out 9.9% worse. One Cursor user saw usage go up, not down. If your own A/B says it costs more, turn it off.
Both shrink tool output before your agent reads it. Run one at a time for a week, so you can tell which one actually helped. Context Mode has its own one-pager.
Anybody can build software now. Not everybody can build a business.
Reading about tools is a great way to feel busy. The Lab is for Founders who'd rather get customers: weekly office hours with a CTO who's built two seven-figure software companies, and a room of other Founders who care whether anybody buys what you're making.
Less Theater. More Customers.
Want my 30-Day Launch Plan? It's for Lab members only: validate, plan, build and launch, one week at a time.
See The Solo Founder Lab ▶