OpenAI has released Codex Micro, a compact physical keypad built with Work Louder for people who spend their day managing Codex agents.
The device is small, niche, and already listed as out of stock on OpenAI's Supply Co page. But it is still worth paying attention to because it gives OpenAI's agent strategy a physical interface: colored status keys for active chats, tactile command buttons for common Codex actions, a joystick for workflows, and a dial for changing reasoning effort.
That makes Codex Micro less like a normal keyboard and more like a control surface for agentic work. OpenAI is not just asking developers to chat with Codex in another window. It is giving power users a way to watch several Codex chats, switch between them, approve or decline requests, dictate prompts, and trigger common workflows without moving fully back into the app.
What Codex Micro is
Codex Micro is a limited-run collaboration between OpenAI's Codex team and Work Louder, the boutique hardware company known for programmable shortcut controllers.
OpenAI prices the device at $230 and lists it as the kbd-1.0-codex-micro. It connects over Bluetooth or USB-C, works with Mac and Windows, and includes:
- 13 mechanical switches
- six frosted Agent Keys for Codex chat status
- one touch sensor
- one rotary encoder
- one planar joystick
- RGB lighting
- a Codex icon keyset with extra keycaps
- Work Louder Input support for deeper customization
The default product page says the device is out of stock, so the launch appears to be a small hardware drop rather than a broad consumer rollout.
What the keys do
The most interesting part is the Agent Keys. OpenAI's documentation says six frosted keys can follow Codex chats and show status through color:
- white for idle
- blue for thinking
- green for complete with an unread update
- amber when Codex needs input
- red for an error
- off when no chat is assigned
By default, those keys follow the six most recently updated chats. Users can also configure them around pinned chats, priority chats, or custom assignments.
That is a very specific interface for a very specific workflow: one person supervising multiple software agents at once. Instead of checking each thread manually, the keyboard makes agent status visible in peripheral vision.
The command surface
Codex Micro also includes command keys for high-frequency actions. OpenAI's default layout includes toggling Fast mode, approving or declining the current request, continuing a chat in a new chat, push-to-talk, and sending the current message.
The push-to-talk key uses the computer's microphone rather than a microphone built into the keypad. Hold the key to record, or double-press it for hands-free recording, then send the prompt when it is ready.
The joystick starts with navigation-style mappings: Plan mode up, forward right, sidebar down, and back left. OpenAI says users can assign available ChatGPT desktop commands or enabled skills to the joystick directions. Work Louder says the device can also use additional layers through Work Louder Input, so it can behave as a broader programmable shortcut pad outside Codex.
The dial is the clearest physical metaphor. It can navigate composer controls, but it can also be set to Reasoning-only mode so turning the dial adjusts reasoning effort. That makes model behavior feel more like a hardware control: low effort for fast changes, higher effort for deeper work.
Why this matters
Codex Micro is not important because most developers need a $230 mini keyboard. They probably do not.
It matters because it shows how OpenAI thinks agent workflows may evolve. The interface problem changes when a user is no longer asking one assistant one question at a time. If a developer is running several agents across code review, debugging, tests, refactors, documentation, and pull requests, the bottleneck becomes supervision.
Codex Micro turns that supervision layer into lights, buttons, and quick controls.
That is a different design direction from the usual chatbot window. It assumes the user is orchestrating parallel work, not simply prompting a single model. The status lights are the strongest signal: OpenAI expects people to care about whether a given agent is thinking, waiting, done, or broken while they are doing something else.
The security tradeoff
There is one obvious caution: physical shortcuts make approvals faster, but faster approvals are not always safer approvals.
Axios noted that an approval button on a physical keypad could make accidental authorization easier. That is a fair concern. Agentic tools often ask for permission before accessing files, running commands, changing code, or taking other actions. A dedicated hardware button reduces friction, which is exactly the point, but it also raises the cost of sloppy habits.
For teams using Codex in sensitive repositories, the right posture is to treat Codex Micro as a convenience layer, not as a substitute for permission discipline. Approval presets, repository controls, review workflows, and human judgment still matter.
Availability
OpenAI's Supply Co listing says Codex Micro costs $230 and is sold while supplies last. The page currently marks it out of stock.
OpenAI's setup guide says orders are expected to begin shipping shortly after purchase, and users configure the device through the ChatGPT desktop app after connecting it by USB-C or Bluetooth. On macOS, ChatGPT needs Input Monitoring permission so it can respond to key presses.
The device is tightly tied to the ChatGPT desktop app and Codex on layer 1. For non-Codex shortcuts and other app workflows, Work Louder Input provides additional layers and custom mappings.
Our take
Codex Micro is easy to dismiss as a collector gadget, and for most users that is probably what it is.
But the product says something useful about where AI work is going. OpenAI is designing around a world where people manage multiple agents, not just type prompts into one chat box. In that world, the problem is not only model intelligence. It is attention, status, control, and trust.
A small keypad will not define that future by itself. Still, Codex Micro is a clear signal: the next interface for AI agents may look less like another browser tab and more like a workstation control panel.