Headroom

Headroom is an open-source tool from Headroom Labs that compresses what an AI agent reads, including tool outputs, logs, retrieval chunks, files and conversation history, before it reaches the LLM. The README's example shrinks a 10,144-token log to 1,260 tokens with the key error line still there. It picks a compressor by content type, with separate ones for JSON, source code and prose, and keeps the originals in a local cache so the model can pull the full text back when it needs it. You can call it inline from Python or TypeScript, run it as a local proxy, wrap a coding agent such as Claude Code, Codex or Cursor with one command, or connect it as an MCP server. What sets it apart is that compression happens on your own machine; the README says no prompt or file content is sent anywhere to be compressed.

Who it's for:developers who lean on coding agents and want to cut token costs. Readers who don't write code won't have a use for it.

Before you start: install it with uv, pip or npm. Compression is local, but the compressed content still goes to your usual model provider, and you pay that provider as before. Original logs and files are cached on your computer. Wrap mode sits between your agent and its model, and headroom unwrap undoes it. headroom learn writes corrections mined from failed sessions into files such as CLAUDE.md or AGENTS.md, so review them first. The README also mentions Headroom for Teams; check the official site for details and pricing. Apache-2.0 licence.

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