Headroom is an 🏷️#open-source context-compression tool from Headroom Labs that shrinks the tool outputs, logs, files, and retrieval chunks fed to large language models before they reach the model. Released under the Apache 2.0 License, it runs locally and reports roughly 20% fewer tokens for coding agents and far larger reductions on structured data such as JSON.
Built in Python and Rust with a TypeScript SDK, Headroom can run as a library, a proxy server, a CLI wrapper, or a 📝Model Context Protocol (MCP) server, and it applies content-aware compressors — separate paths for JSON, code, and prose — whose output is reversible through a local cache. It also maintains a cross-agent memory store shared across assistants such as Claude, Codex, Gemini, and Grok, and integrates directly with coding agents including Claude Code, Cursor, and Copilot.
Key Features
- Content-aware compression — Separate compressors for JSON, code, and prose, each tuned to its format rather than a single generic pass.
- Reversible output — Compressed content is cached locally and can be retrieved in full when the original is needed.
- Flexible deployment — Runs as a library, proxy server, CLI wrapper, or MCP server.
- Cross-agent memory — A shared context store spanning Claude, Codex, Gemini, and Grok.
- Local processing — Compression happens on-device; no prompts or files are sent out to compress them.