shubhamsinghrathore/headroom ? reverse-engineered prompt

Build me a local first tool called Headroom that sits between an AI coding assistant and the model, and shrinks whatever gets sent over, things like tool output, logs, files, retrieved docs, and chat history, so the model uses way fewer tokens but still gives basically the same answer. I want it to work a few ways, as a simple library I can call from code, as a proxy I can point apps at without changing much, and as a wrapper around common coding agents so I can turn it on quickly.

It should be reversible, so the original content is cached locally and can be fetched if the model really needs it. Add a small MCP server with tools to compress, retrieve original content, and show savings stats. Include a health check, a live dashboard, and some memory across sessions so different agents can share context and avoid duplicates. If possible, also trim overly wordy model responses. Keep the experience simple, safe, and local. Look up current docs online if you need to.