Headroom: Cut Your AI Agent's Tokens by 90% (Open Source)

Every modern AI agent drowns in its own tool outputs — giant JSON blobs, thousand-line logs, whole files. The model only needs the signal, but you pay for all the noise, every token, every turn. Headroom fixes that: an open-source context-compression layer that sits between your agent and the model and shrinks everything the model reads before it's ever sent.

In the video I install it and watch it intercept live Claude Code traffic — compressing one request from 3,576 → 666 tokens (81%). The community running it has already saved 59 billion tokens (~$200,000 in API bills).

What it does

• 60–95% fewer tokens, with answers and accuracy held
• Three ways to drop it in: Python/TypeScript library (one compress call), a proxy with zero code changes, or an MCP server for any client
• Smart Crusher for JSON arrays, plus code, images, and logs
• Cuts output tokens too (which cost ~5× input on a model like Opus)
• CCR — reversible compression. Local-first.

One command wraps your coding agent: headroom wrap claude.

Links

• GitHub → https://github.com/chopratejas/headroom
• Docs → https://headroom-docs.vercel.app/docs
• Model (Kompress-v2-base) → https://huggingface.co/chopratejas/kompress-v2-base

Full walkthrough + live demo in the video above. New videos on local LLMs, AI agents, and automation every week — subscribe.