headroom

Why we included this project

Headroom cuts the token bill before your prompt ever reaches the model, and it does it without making you change how you call the API. You can run it as a proxy, import it as a Python or TypeScript library, or wire it into a framework, and it figures out on its own whether a payload is JSON, code, logs, diffs, or plain text so it can pick the right compressor. The savings are real: around 20% fewer tokens for coding agents and 60-95% fewer for JSON-heavy payloads. It only compresses the newest content blocks, leaves the prompt-cache hot zone alone, and keeps the originals around so the model can pull back full detail when it needs it. Teams on Claude Code, Cursor, or LangChain/Agno agents with large codebases or long debugging sessions will see the savings add up, and the proxy path needs zero code changes, so you can test it without touching your stack.