Compresso runs Headroom for you.

What each one actually gives you

Both columns run the same engine and compress the same way, on your own machine. The differences are all in the work around it.

	Compresso	Headroom, yourself
Install	Download the app and open it. Nothing is added to your PATH.	pip install into a virtualenv you create and maintain.
Python	A pinned interpreter and a managed virtualenv, private to the app.	Yours to provide, and yours to keep working.
Engine version	Pinned per release and integrity-checked. A checksum mismatch fails the launch rather than running anyway.	Whatever you install or pin yourself.
Pointing an agent at it	One click per agent. Disconnecting restores the original file exactly.	Edit each agent's own config, and remember what you changed.
Keeping it running	A background helper, with a menu bar item and a pause switch.	You run the proxy and keep it up.
Seeing the savings	Per agent and per day, with removed tokens counted separately from provider cache discounts.	Whatever the engine reports on your terminal.
Platform	macOS 14 and later, Apple Silicon only.	Anywhere Python runs: Linux, Windows, macOS.
Surfaces	The proxy path, for Claude Code and Codex.	Proxy, SDK, CLI wrapper and MCP server.

Provisioning, pinning, and staying out of the way

The engine is a Python package, and a Python package is a thing you have to install, isolate, version and keep alive. Compresso does that part: a private interpreter and virtualenv it manages itself, a version pinned per release, downloads checked against a hash where a mismatch is a hard failure rather than a warning, and a supervised process you can pause from the menu bar.

The rest is the part you would otherwise be doing by hand. Compresso edits each agent's own config to point it at the local intercept, keeps a timestamped backup of what it changed, and puts the file back exactly as it was when you disconnect. Nothing you use ends up depending on Compresso still being installed.

The privacy story is unchanged by any of it, because the engine already worked this way. Compression happens on your Mac. Your prompts and your code go straight from your agent to Anthropic or OpenAI, and never to us. That is covered in more detail on Trust & security.

For what the pipeline does to your context, see How it works. For the numbers, including the workloads where compression saves nothing at all, see Benchmarks, which cites the engine's own published runs rather than any measurement of ours.

Credit where it belongs

Headroom is developed by Headroom Labs and licensed under Apache-2.0. The source is on GitHub, and you are free to run it yourself without Compresso or any other app. We are not affiliated with Headroom Labs, and this page is not endorsed by them; "Headroom" is used here to name the software Compresso runs. The engine keeps its own license and its own authors, and downloading Compresso grants you no rights over either.