The problem it solves
Coding agents do not mostly send your prompts. They send context: the files they read, the output of every command they ran, search results, documentation, and the entire previous conversation, re-sent on every turn. Your own typing is a rounding error next to it. That machine-generated bulk is what drains a 5-hour window by lunchtime, and it is highly compressible, because most of it is repetitive, structured, or simply not needed again.
More detail on where it goes: why Claude Code is so expensive, or the same for Codex.
How Headroom works
Headroom runs a local proxy on your machine and points Claude Code and Codex at it. Before a request goes to the model, Headroom rewrites the bulky, repetitive parts of the context into a compact form, then forwards the leaner version. Nothing is deleted: it attaches a retrieval tool, so if the model later needs the full detail of something that was compressed, it asks and gets the original back.
That reversibility is the part that makes it safe to use on real work. Instead of re-paying for a full log or document on every single turn, you carry a compressed version and expand the occasional slice once, which is where the net saving comes from. It is also designed to preserve the cached prefix of your conversation, so it does not sabotage prompt caching to win on raw token count.
Because it sits on the request path rather than inside the agent, it covers every session automatically, including sub-agent runs. There is no plugin to install and nothing to invoke.
Does it hurt quality?
Not measurably, on the open-source benchmark suite. Compressed context scores 0.919 F1 with 98.2% recall on HTML extraction, retrieves 4/4 needles from JSON log haystacks, and slightly improves QA F1 on SQuAD v2 and HotpotQA. Median overhead is roughly 52 milliseconds per request, which is about one network round-trip and invisible next to the seconds the model itself takes. And when in doubt, the model can always pull the original content back.
Where your code goes
Nowhere. Compression happens locally, on your machine, which is the whole reason the design works: your prompts, code, and files never need to reach a Headroom server to be optimized. What does leave is aggregate savings statistics tied to your account. The privacy policy spells out exactly what that covers.
Is Headroom free?
The compression engine is the open-source Headroom CLI, free to self-install and run from the terminal. The macOS app is built on that engine, with the endorsement of the CLI's maintainer, and adds the one-click install, signed and notarized builds, auto-updates, the menu bar UI and savings stats, and the add-ons screen.
The app starts with a 7-day free trial, no credit card required. After that, plans match your Claude or ChatGPT tier and are priced at a fraction of the subscription they stretch. See the pricing section for current numbers, or app vs CLI if you are weighing the free path.
Is it safe to install, and easy to undo?
It is a signed, notarized macOS app. It does not modify Claude Code or Codex themselves; it routes them through a local proxy by setting a base URL and a few managed configuration blocks. Quit Headroom from the menu bar and those revert immediately, so your agents talk to their APIs exactly as they did before. Uninstalling is dragging the app to the Trash, and that page lists the few leftovers if you want a clean sweep.
What else is in the box
Compression is the core, but three one-click add-ons cover waste it does not reach. Ponytail is a skill that makes the agent write less code, attacking the output side of the bill. RTK trims noisy terminal command output before it is ever recorded. MarkItDown converts PDFs and Office files into compact Markdown. The app also tracks your usage windows, so you can see how much of your 5-hour and weekly limits are left.
Who it is for
Developers who lean on Claude Code or Codex daily and keep running into limits: hitting the 5-hour window mid-afternoon, burning the weekly cap by Wednesday, or watching a large-codebase session eat its budget on file reads. If you are a light user who never hits a limit, you do not need this. If you are upgrading tiers to buy headroom, that is exactly the spend it is meant to replace.
More questions are answered on the FAQ.