Headroom - Compress 60-95 Percent of AI Agent Tokens Without Losing Quality

The Token Problem

Every AI agent has a hidden tax. Tool outputs logs RAG chunks files conversation history all consume tokens. Headroom a new open-source project currently #1 on GitHub Trending with 3500 stars today solves this.

How It Works

Headroom compresses everything your AI agent reads before it reaches the LLM. Tool outputs logs RAG chunks files conversation history. Same answers fraction of the tokens.

Live example from their README. 10144 tokens compressed to 1260 tokens and the same answer was found.

Four Ways to Use It

1. Library - compress in Python or TypeScript inline in any app
2. Proxy - headroom proxy port 8787 zero code changes
3. Agent Wrap - headroom wrap claude codex cursor aider copilot
4. MCP Server - for any MCP client

What Makes It Special

Reversible Compression CCR. Originals never deleted LLM retrieves on demand. Cross-Agent Memory shared store across Claude Codex Gemini with auto-dedup. Self-Learning headroom learn mines failed sessions writes corrections to CLAUDE.md AGENTS.md.

Why This Matters

AI agents need context. Context costs tokens. Headroom lets agents keep more context for less cost. Essential for workflows with long tool execution chains.

pip install headroom
headroom wrap claude

Or point your app to headroom proxy port 8787.