Headroom

Headroom sits between your AI coding agent and the tokens it burns. It compresses the noisy input that Claude Code and Codex swallow on every run, and it writes down what your agent learns about each project so it stops paying to rediscover the same things tomorrow.

What Headroom actually does

The app runs on your own machine, on Mac or Windows, and works on two fronts: shrinking what goes into the model, and giving the agent a memory so less needs to go in at all.

• Track total token costs and input savings over time on a savings dashboard
• Chart daily savings, split between compression and output savings
• Capture reusable commands and environment details per project
• Auto-write learnings to CLAUDE.local.md and AGENTS.md when a pattern is identified
• Review compressions, rescan reminders and large savings events in a live feed
• See cache hits and compression broken out, with a plain explanation of the math

The features that matter

Most token tools stop at compression. The part worth paying attention to here is the learning loop. Every time your agent works out how your repo is laid out, which test command actually runs, or which environment variable it keeps tripping over, Headroom writes that into the project files the agent already reads at start-up. The next session begins with the answer instead of twenty tool calls spent finding it. Compression saves you tokens on a single run; memory saves them on every run after that.

The second piece is the add-on manager. Headroom can switch on Ponytail, Caveman, RTK, MarkItDown, Serena, Codebase Memory or Context7 from inside the app, with one click each. Those tools go after what plain compression misses: terminal noise, binary documents, whole-file code reads and replies that ramble. You could install and wire each of them up yourself, and plenty of people do. Having them in one panel, running locally and kept away from your project dependencies, is the convenience you are buying.

Everything runs on your machine. Nothing from your codebase is shipped off to a third-party server for the compression step, which matters if you work on client code under an NDA. The compression is also reversible, so a trimmed input can be restored rather than lost.

Who it's built for

This is for developers who already live inside Claude Code or Codex and keep hitting usage limits before the week is out. If you pay for Claude Pro or ChatGPT Plus and feel the ceiling every few days, the vendor's claim is that you get up to double your available coding usage. Heavier users on Claude Max or ChatGPT Pro plans are covered by the higher tiers.

If you use an AI agent twice a week for a quick script, skip it. The savings compound with volume, and a light user won't see enough of them to care. My take: the compression is the headline, but the project learnings are the reason to keep it installed. An agent that remembers your repo is simply faster to work with, whatever it saves on the bill.

How you actually use it

• Install the signed desktop app on your Mac or Windows machine.
• Keep working in Claude Code or Codex the way you already do.
• Open the savings dashboard to see what was trimmed and what was learned.
• Turn on the add-ons that match where your sessions still waste tokens.

The deal

The entry tier is $29 (was $72) for one user, and it works with Claude Pro and ChatGPT Plus. Tier 2 at $99 adds Claude Max x5 and ChatGPT Pro Lite, and Tier 3 at $199 covers Claude Max x20 and ChatGPT Pro. Pick the tier that matches the AI plan you already pay for, since that is what decides how much there is to save.