Skip to main content
ClaudeREF.com
← ecosystem index
output styleverified 3 August 2026

Caveman

Julius Brussee

Coding agents often spend many tokens explaining what they are doing, even when a short answer would be enough. Caveman changes the agent's writing style so replies are compressed while code, paths, commands, and technical details stay intact. Shorter replies can reduce output-token usage, cost less to read back into a long conversation, and make the session easier to scan. It mainly changes how much the agent says; it does not make the underlying model smaller or guarantee the same saving on every coding task.

Best for

People who want concise agent replies, lower conversational token usage, and less text between code changes.

License

MIT

Last pushed

4 August 2026

Compatibility

Claude Code · Codex · Gemini CLI · Cursor

Why it is interesting

It includes several compression levels, session activation, token-usage statistics, and tools for shortening memory files and MCP descriptions.

Check before installing

The repository's benchmark measures specific tasks and configurations, so treat its percentage savings as examples rather than a guarantee. Use a less aggressive mode when security explanations or destructive actions need more context.

Documented install

claude plugin marketplace add JuliusBrussee/cavemanclaude plugin install caveman@caveman