Every agent session burns tokens on stuff the model barely needs. A 2,000-line log. A huge JSON response from an API. A whole file read just to find one function.
What Headroom does
I came across Headroom this week. It sits between your agent and the model and compresses tool outputs before they reach it: logs, JSON, diffs, file reads, search results.
The bit that caught my eye is that the compression is reversible. The original goes into a cache, and the model gets a tool to pull the full version back if it needs it. Nothing is thrown away, it just doesn't get sent by default.
It works with Claude Code as a proxy with no code changes, or through MCP.
The numbers
The headline says 60–95% fewer tokens, but that's for JSON. For coding agents they say around 20%, and their own benchmarks range from 21% on code search to 57% on incident debugging. I trust those smaller numbers more, and they're still worth having.
Why I care
I haven't tried it yet. But my Claude stats show 29.6 billion tokens, and most of that is agents reading things, not writing them. So this one's going on my list.
If it holds up, I'll share real numbers from my own setup.
Top comments (0)