Caveman is a token-saving toolkit for AI coding agents. It makes your agent write terser responses and compresses the logs, test output, JSON, diffs, and web pages it reads—helping reduce token usage without replacing code, commands, file paths, or error messages.
Use the lightweight skill to make supported agents reply in a compact “caveman” style, or run the local proxy to shrink the large tool outputs agents repeatedly send to model providers. For developers building their own AI applications, the middleware applies the same compression around existing OpenAI, Anthropic, Vercel AI SDK, LangChain, and other provider calls.
• Features
Token-saving agent skill — Shortens agent prose while preserving code, commands, file paths, and exact error messages
Local compression proxy — Compresses logs, test output, JSON, diffs, search results, and browser content before they reach the provider
Reversible compression — Stores byte-exact originals locally in SQLite and provides recovery handles
Works with many agents — Supports Claude Code, Codex, Gemini CLI, Cursor, Windsurf, Cline, Copilot, Aider, Qwen Code, OpenCode, and more
Developer middleware — Integrates with TypeScript and Python agent frameworks through a small wrapper
Token analysis — caveman learn reviews local agent history, identifies token-heavy configuration and context, and suggests fixes
Useful agent commands — Includes terse commit messages, one-line code reviews, memory-file compression, usage statistics, and compressed subagent presets
Browser compression — Provides a compressed page view for agents instead of large accessibility snapshots
Privacy-focused operation — The skill runs locally and requires no account or API key; proxy originals remain on your machine
Open source — Licensed under Apache-2.0
• How it works
Skill: Adds a rule file that changes how the agent phrases answers
Proxy: Runs locally between your agent and provider, compressing readable context before each request
Middleware: Compresses tool results inside your own application while retaining originals in conversation history
No comments or reviews, maybe you want to be first?