Caveman

Caveman is an open-source set of token-saving tools for AI coding by Julius Brussee (GitHub user JuliusBrussee), and its README opens by saying it makes your AI agent say less and read less. The full version is a command-line tool with six parts once installed: a voice Skill for shorter answers, a local proxy that shrinks logs and JSON before the model reads them, Auto, which picks the model and effort for each ask, Blocks, which lets the agent reuse scripts it already wrote, waste fixes that show where your tokens go, and compressed web browsing. AI Inspiration World recommends only the "say less" part, the caveman Skill. With it on, the agent gives the answer first, then the reason and the next step, with no greetings, no recaps and no "hope this helps", one idea per sentence, following rules based on ASD-STE100, the simplified English used in aircraft maintenance manuals. /ultracave goes shorter still, /megacave answers in classical Chinese, and saying "stop caveman" switches it off. What sets it apart is that it compresses only the voice: code, commands, paths, numbers and error messages stay character for character, words like not, never and only are never dropped, and for security warnings, irreversible actions or a confused user it switches back to full sentences. The README also shows the numbers that cut against it: on the author's ten developer questions, compared with simply asking for concise answers, /caveman saves only about 3% more at the median and /ultracave about 35% more; the rules themselves cost about 1,160 input tokens, and on tools billed per request (the README names GitHub Copilot) shorter answers save nothing.

Who it's for:developers using AI coding tools such as Claude Code or Codex who find the answers too wordy and want less to read. This is a developer tool; readers who don't write code won't get much from it.

Before you start: we suggest copying only the skills/caveman folder into the folder your AI tool reads Skills from; /caveman ultra and /caveman wenyan hand off to ultracave and megacave, so copy those two folders as well if you want them. The README's command for the Skill alone, without the proxy, is npx skills add JuliusBrussee/caveman -g, which downloads a third-party installer through npx, and we haven't checked which of the repo's Skills it installs. SECURITY.md says the Skill runs on your machine and doesn't contact Caveman's servers. We have not reviewed or recommended the full command-line tool the README leads with, npx @caveman-ai/cli (Node.js 22.13 or later). It finds the AI tools and logins on your computer, installs hooks and a status line, points the request address of tools such as Claude Code (ANTHROPIC_BASE_URL) at a local Caveman proxy, and turns on all six parts after install. The local proxy handles your requests and model credentials, and it keeps the original, uncompressed content in ~/.caveman/ccr.db on your machine, which can include prompts, tool output and credentials embedded in that content; SECURITY.md says the file is not encrypted and uninstalling doesn't delete it. The CLI sends usage statistics by default: commands run, token counts, a random install ID, your OS and your IP address (IPs are kept for 90 days), never prompts or code; caveman telemetry off or DO_NOT_TRACK=1 turns it off for good. Auto needs a Caveman account through caveman login. Once you pick Auto, every ask sends Caveman Cloud your latest message (with whatever your agent attaches, such as CLAUDE.md, memory, @-mentioned files and command output), the message before it and the end of the agent's last reply, unedited, and on the free plan Caveman may keep this to improve routing; requests routed to a Caveman Cloud model pass through Caveman Cloud whole, including the system prompt, tool results and code. Keys you add for other model providers are stored in the macOS Keychain or a local file. We have not reviewed or recommended the local proxy, the engine, Blocks, the MCP compressor, the browsing feature that can navigate and click web pages, the Caveman Cloud service, or the README's install commands that pipe a remote script straight into a shell (curl ... | bash, irm ... | iex). Apache-2.0 license; "Caveman" and the rock logo are trademarks of the author.

Based on the project's public documentation. Not tested by AI Inspiration World.