August 2026
A second savings layer, honest metrics, and a Windows preview
Headroom found a second thing worth compressing. Besides tool output, your coding agent re-sends its full tool definitions - the JSON schemas describing every tool it can call - with every single request, unchanged, thousands of tokens at a time. Headroom now compresses those too, automatically, and on busy sessions this layer is often the largest share of what it saves.
The savings numbers themselves also got a rework. The headline rate is now the share of non-cache-read input removed - the input you would actually have bought this turn - with cache reads shown as context instead of claimed as wins, plus a separately labeled, live-sampled estimate for output-side savings. For the full arithmetic on a real day of traffic, how to present token savings percentages honestly is the long answer; these numbers are the honest ones.
Beyond the numbers: an early Windows preview now ships alongside every stable release as an x64 installer on the releases page, marked experimental in-app while it hardens. macOS users can install with brew install --cask headroom. And teams got their first admin feature: a per-person savings breakdown, so you can see the total across every seat.
July 2026
OpenCode and Grok Build connectors
Two more coding agents join Claude Code and Codex. Enable OpenCode and Headroom routes it through the same local optimization proxy - across every provider OpenCode can talk to, not just one - via a transport plugin Headroom installs for you. Grok Build is routed the same way to xAI's API. Both are experimental, and both are the usual one-click setup.
The rest of the app treats them as first-class citizens: Project learnings scans past OpenCode and Grok sessions and writes token-saving patterns into each agent's own memory files, and the menu bar home view gained compact connector badges so you can see at a glance which agents are routed and healthy.
June 2026
Codex support and one-click add-ons
Headroom now supports OpenAI Codex as a first-class connector, not just Claude Code. Enable Codex in the app and Headroom routes the Codex CLI through the same local optimization proxy, reversibly compressing tool output and boilerplate before it reaches the model. The same prompt fits in fewer tokens, so the ChatGPT plan you already pay for lasts longer. Headroom also tracks your Codex usage windows, so you can see how much of your 5-hour weekly limit you have left right from the menu bar.
The same release adds an Add-ons screen: a small set of optional tools you can install and toggle with one click, wired up for both Claude Code and Codex with no separate setup.
- RTK is a token-optimized shell proxy that trims noisy terminal output, so your agent sees the important parts of a command's results without the clutter.
- MarkItDown converts PDF and Office documents to Markdown so they cost far fewer tokens when your agent reads them.
- Ponytail nudges the agent to write the least code possible, attacking the output side of the bill that input-focused tools leave untouched.
If you keep hitting the cap, the Codex usage limits guide covers it, and how to reduce Codex costs walks through the rest of the stack.
May 2026
Activity feed and Project learnings
Two features made Headroom less of a black box. The Activity feed gives you one place to see what Headroom has been doing: savings recorded, available updates, and warnings as you approach a usage limit, so a hit cap never comes as a surprise.
Project learnings turns repeated mistakes into permanent fixes. Headroom scans your past Claude Code and Codex sessions and writes token-saving patterns into each agent's own memory: CLAUDE.md and MEMORY.md for Claude Code, AGENTS.md and instructions.md for Codex. When an agent repeats a mistake, Headroom updates that memory so it doesn't happen again, and you can ask it to scan history on demand to seed those learnings yourself.
April 2026
Headroom launches for Claude Code
The first version of Headroom did one thing: cut Claude Code token costs without changing your workflow. It runs as a menu bar app that routes Claude Code through a local proxy and reversibly compresses tool output, boilerplate, and large inputs before they enter the conversation. When the model needs something that was compressed, a small retrieval tool pulls the original back on demand, so nothing is thrown away.
Because optimization runs locally on your machine, your prompts and code never need to leave it. The app tracks token and dollar savings over time, and the compression engine is the open-source Headroom CLI, packaged into a signed, auto-updating macOS app. Codex support and the add-ons above came later, but this is where it started.