Headroom.

Headroom is a macOS menu bar app that cuts the token consumption of Claude Code and OpenAI Codex by roughly 50%— which means the subscription you already pay for lasts about twice as long before you hit usage limits. It sits between your coding agent and the model as a local proxy, quietly compressing the noise out of every prompt before it's sent.

The desktop app is built by Garm Tech on top of the open-source Headroom CLI created by Tejas Chopra — a project that picked up tens of thousands of GitHub stars within months of release. The app wraps that engine in a zero-setup package: install it once and it optimises every session automatically, with a dashboard showing exactly how many tokens (and dollars) it saved you.

Headroom runs as a local proxy: it intercepts each request before it leaves your machine and reversibly compressesthe tool output, logs, and boilerplate that bloat it. "Reversible" is the key word — nothing is permanently thrown away. A small retrieval mechanism lets the model pull back the original, uncompressed content whenever it actually needs a detail, so you only pay full price for the parts that matter.

Three things are worth knowing about the trade-offs:

• Privacy.All optimisation runs on your machine. Your prompts and code never touch Headroom's servers, and the app is a signed, notarised macOS binary.
• Quality. Because originals stay retrievable, answer quality holds up — the team reports benchmark results like 0.919 F1 with 98.2% recall on HTML extraction tasks.
• Escape hatch. Disable or uninstall it and your tools instantly go back to talking to the model directly. Nothing about your setup is permanently changed.

There are two ways to run Headroom, and choosing is mostly a question of how much you want to think about it:

Both deliver roughly the same ~50% token reduction — the engine is the same. If you're comfortable in a terminal and happy to babysit a process, the CLI costs nothing. If you just want the savings to happen, the app is the "set it and forget it" option.

Setup is deliberately boring — there's no configuration step at all:

• Download the macOS app from extraheadroom.com.
• Drag it into your Applications folder.
• Open it once. It configures itself, parks in your menu bar, and starts optimising every Claude Code and Codex session in the background.

Enter your email and a 7-day free trial starts — no credit card required. During the trial you get full optimisation; when it ends, you pick the plan matching your Claude or ChatGPT tier. The menu bar dashboard shows compressions as they happen, tokens saved, and a weekly recap.

Prefer free and open source? The CLI needs Python 3.10+ and a little terminal comfort. Install it with:

Then wrap your coding agent so every session runs through the compressor:

There are two other modes if wrapping doesn't fit your workflow — a standalone proxy you point tools at, and an MCP server registration:

The app ships with an Add-ons screen — optional tools that attack token waste from other angles, each toggled with one click:

• RTK — trims noisy terminal output before it reaches the model. Build logs and test runs are some of the worst token offenders.
• MarkItDown — converts PDF and Office documents into Markdown, so the agent reads clean text instead of choking on raw file formats.
• Ponytail — nudges the agent to write the least code possible, which saves tokens on the output side and tends to produce tighter diffs anyway.

Plans are priced against the AI subscription they're stretching, at founder pricing (40% off) at the time of writing:

• Pro — from ~$3/month, for Claude Pro or ChatGPT Plus users.
• Max ×5 — from ~$12/month.
• Max ×20 — from ~$24/month.
• Team / Enterprise — custom pricing.

Annual billing saves a further 33%. The pitch is simple arithmetic: if a few dollars a month roughly doubles the usage you get from a $20–$200 subscription, the tool pays for itself several times over — and the dashboard shows you the actual savings so you can check the maths on your own usage before the trial ends.