GitHub radar
Headroom: Reduce AI Agent Tokens By 60–95%
A context compression layer between your AI coding agent and LLM that compresses tool output, logs, files, history etc. before sending to the model. Get the same answers using less tokens!
Headroom is a context compression layer developed by Headroom Labs. It sits between your AI agent and LLM and compresses tool output, logs, RAG fragments, files, history etc. before sending to the model. It compresses JSON 60–95% and code ~20%, but the answers are the same as if they were not compressed. It can be attached to your existing agent in a single command (headroom wrap claude), acts as a transparent proxy listening on port 8787 without modifying any code, is available as a Python/TypeScript library, and can be run as an MCP server. Works with Claude Code (Anthropic), Codex (OpenAI), Cursor, Grok, Copilot, Cline, Goose, OpenHands, and more. Shared memory store for multiple agents allows consistent context between Claude, Codex, Gemini, Grok, etc. headroom learn analyzes previous failed sessions and stores corrections in CLAUDE.md to prevent future errors.
Why a vibe-coder should care
This can save you real money if you are using Claude Code or another AI agent every day, because it compresses all the repetitive tasks (like reading files, parsing logs, working with JSON) before sending them to the expensive LLM. This compression happens locally, without sharing any information with external servers.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Install headroom from https://github.com/headroomlabs-ai/headroom and wrap my Claude Code with it — follow the README, run `headroom wrap claude`, then show me the token compression stats. Ask me if you need permissions.
Runs on a regular laptop — only Python or Node needed. No GPU required.
Open on GitHub▌ More finds