In One Sentence
Caveman is an open-source Claude Code skill plugin whose core idea is brutally direct — make the AI talk like a caveman, deleting every pleasantry and explanatory windup, cutting output tokens by roughly 75 percent while keeping technical content complete and accurate. The project slogan: “why use many token when few do trick.”
The Project at a Glance
| Project | Caveman |
| Author | Julius Brussee |
| GitHub | github.com/JuliusBrussee/caveman |
| Core function | Cuts AI output tokens ~65-75% while preserving technical precision |
| Supported platforms | Claude Code, Gemini CLI, Codex, Cursor, Copilot, and 30+ other agents |
| License | MIT, fully open source |
Why Caveman Exists
Anyone who has used Claude Code or a similar AI coding agent knows the pattern: the AI pads every answer with politeness, explanatory preambles, and repeated confirmations. Ask “why does this React component keep re-rendering” and the AI may spend dozens of tokens warming up before the one sentence that actually answers the question.
Those filler words are harmless in casual chat. For developers hitting APIs at high frequency, every wave of output is real money. Caveman aims squarely at this pain point: the shortest sentence carrying the most information.
The Effect, Side by Side
Normal mode (69 tokens):
“The reason your React component is re-rendering is likely because you are creating a new object reference on each render cycle. I would recommend using useMemo to memoize the object.”
Caveman mode (19 tokens):
“New object ref each render. Inline object prop = new ref = re-render. Wrap in useMemo.”
Identical information. 72 percent fewer tokens.
Multiple Intensity Levels
Caveman is not one-size-fits-all; it offers several “caveman grades”:
- Lite: keeps basic sentence structure, removes pleasantries and redundant explanation
- Full: compresses further, replacing full sentences with arrows and keywords
- Ultra: maximum compression — only the core technical information survives
- Classical-Chinese (wenyan) mode: outputs in compressed literary Chinese style for Chinese-language scenarios — the same re-render answer rendered as terse classical prose
The Agent Ecosystem It Supports
Caveman started as a Claude Code plugin but has grown into a cross-platform ecosystem supporting more than 30 AI coding agents: Claude Code (native plugin install), Gemini CLI (extension install), OpenAI Codex, Cursor and Windsurf (IDE rules files), Cline and GitHub Copilot, plus 20+ more including opencode, Roo, Amp, Goose, Kilo, Devin, Tabnine, Trae, and Warp.
One-Line Install
Installation is a single command:
# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.sh | bash
# Windows PowerShell
irm https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.ps1 | iexThe script auto-detects which agents are installed on your machine and configures Caveman mode for each one. To target a specific agent only:
curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.sh | bash -s -- --only claude-codeThe Bonus Tools
Beyond the core conversation-compression mode, the project bundles several practical extras:
- Caveman Commit: generates minimalist git commit messages — “fix login bug,” “add export btn”
- Caveman Review: one-line code reviews that name only the core problem
- Caveman Compress: an input-compression tool that cuts input tokens by about 46 percent per session
- Caveman Stats: tallies exactly how many tokens you have saved
What Caveman Says About Where AI Tooling Is Going
It is worth reading Caveman as more than a money-saving hack, because the enthusiasm around it tracks a genuine shift in how models are used. Politeness, hedging, and restatement are artifacts of models trained to converse with humans — the register that made ChatGPT feel magical in a chat window. But an agent’s output is increasingly consumed by other programs: its reasoning becomes another model’s input, its plan becomes a scheduler’s instruction list, its answer becomes a CI pipeline’s next step. In machine-to-machine traffic, every word of human-warmth is pure overhead, and compression is not a stylistic preference but an engineering discipline. That is why the plugin’s adoption curve looks the way it does — the same logic pushing OpenAI to compress reasoning traces and enterprises to set token quotas is pushing individual developers to install a plugin that makes their agent say less. Caveman’s four intensity levels matter for the same reason: compression is a dial, not a switch. Lite for code review where some context helps; Ultra for mechanical edits where only the instruction matters. The developers getting the most out of it treat verbosity itself as a cost line item — which, in 2026, it simply is.
The Bottom Line
Caveman looks like a joke and works like a lever. One simple creative move — make the AI speak like a caveman — addresses the developer’s real pain of managing API token costs head-on. For heavy users of AI coding agents, saving 65-75 percent of output tokens means the same budget does more work. Better still, it has grown into a cross-platform skill ecosystem covering nearly every mainstream AI coding tool, MIT-licensed so anyone can use and modify it freely. If you are a heavy AI coding user, Caveman is worth an install. After all — why use many token when few do trick.
One practical note for teams: because Caveman changes the agent’s output register, it also changes what your logs and audit trails look like. The terse mode is wonderful for your terminal and your invoice, and slightly harder for a colleague skimming session history to follow. The sensible pattern many adopters land on is intensity matching — Lite or Full for interactive work where humans read along, Ultra reserved for scripted flows where no human is watching and every token is pure freight cost.
