Contents
Artificial Intelligence
On-premise AI architectures, local LLMs, RAG, autonomous agents. Intelligence you cannot afford to see switched off must be owned.
Explore →
Admina Enterprise
Open Source AI governance: audit trail, PII redaction, bidirectional ALLOW/BLOCK/REDACT policies on any model, local or remote.
Explore Admina →
Research & Development
Applied research and prototyping, from AI to infrastructure, in collaboration with universities and research centres.
Explore →
Linux Services & Systems
Design, deployment and management of on-premise Linux infrastructure: from servers to AI hardware, with data that stays yours.
Explore →A keyboard to command the agents
On July 15, 2026 OpenAI opened orders for its first branded hardware: the Codex Micro, a small $230 desktop macropad, designed with Work Louder (a boutique keyboard maker) and sold under the “Supply Co.” label as a “Co-Lab” collaboration. It is not a computer, not a phone: it is a control surface for people who use Codex, OpenAI’s coding agent, and who juggle several threads of it at once.
The object is concrete and a little playful. Six backlit “agent keys” signal each Codex thread’s status by colour: white idle, blue working, green complete, amber when it needs your input, red on error. There are customisable command keys as shortcuts for frequent actions, a joystick to launch workflows (review a PR, debug, refactor) and a dial that adjusts how much “reasoning”, meaning time and compute, an agent spends on a task. It is configured and driven from the ChatGPT desktop app. It is a limited run, aimed at the most intensive users among the roughly 5 million who use Codex each week, with first deliveries estimated around July 24.
Note. The facts, figures and dates in this article come from official announcements and the trade press, as of mid-July 2026. Some of the products mentioned (the Codex Micro itself, OpenAI’s device with Jony Ive) were not yet in users’ hands: where that is the case we say so, and we do not attribute market outcomes that do not exist yet.
It is worth reading for what it is: not the new frontier of devices, but a well-made accessory for a narrow audience. Fittingly, the launch came weeks after OpenAI’s head of Applications, Fidji Simo, told staff not to be distracted by “side quests”.
The big device, the one you cannot see
The Codex Micro is not the Jony Ive device. That one, born from the roughly $6.5 billion acquisition of io (announced May 2025), remains invisible. The “io” name was dropped after a trademark injunction won by the startup iyO; court filings reported in February 2026 put it at no earlier than end of February 2027, slipping from an original goal of end-2026; and on July 10, 2026 Apple sued OpenAI for alleged trade-secret theft on the hardware side, with OpenAI replying on July 14 that it saw no evidence the complaint had merit. Descriptions speak of a screenless AI companion for the home, with a camera and sensors; a price around $200-300 remains an unconfirmed rumour.
The contrast is instructive. On one side a $230 macropad shipping in days; on the other the “third device” to sit beside your phone and laptop, announced with fanfare, still stuck between delays and lawsuits. It is the difference between augmenting a tool that exists and promising a brand-new device. And it is exactly the line that, over the past two years, has separated the AI hardware that works from the AI hardware that does not.
The graveyard of “AI computers”
The list of attempts to replace the phone with a smart box is by now a small graveyard.
The Rabbit R1 (January 2024, $199, designed by Teenage Engineering) promised a “Large Action Model” that would use apps on your behalf. Reviews were brutal (Marques Brownlee called it “barely reviewable”), the software turned out to run as an ordinary Android app on a Pixel, and in June 2024 a group of researchers found API keys hardcoded in plain text. In mid-2024 CEO Jesse Lyu claimed about 130,000 units sold; by September the daily active users were around 5,000. Rabbit kept pushing software (an agent, a new OS) and in March 2026 previewed a second device, “Project Cyberdeck”: for now a teaser, with no final specs and no date.
The Humane AI Pin is the textbook case. Founded by two ex-Apple designers, Humane raised about $230 million and in November 2023 launched a screenless pin that projected its interface onto your palm with a green laser, at $699 plus $24 a month. In practice: overheating, a few hours of battery, slow and often wrong AI. Brownlee called it “the worst product I’ve ever reviewed”. By the summer of 2024 returns were outpacing sales. In February 2025 HP bought its assets for $116 million, and on February 28, 2025 every Pin was remotely disabled, all at once. The technology survives only as software: in March 2026 HP announced “HP IQ”, an on-device assistant for its business laptops.
And Friend, Avi Schiffmann’s “always-listening” pendant, who paid $1.89 million for the friend.com domain. Announced in July 2024 at $99, shipped in the summer of 2025 at about $129, it ran on Claude 3.5. Its New York subway campaign (over a million dollars, September 2025) was covered in scrawls reading “AI is not your friend”; Schiffmann replied that “capitalism is the greatest artistic medium”. Sales of a few thousand units, and a spot in the Museum of Failure. In the same space the open-source rival Omi ($89) carries on, more quietly.
Who actually sells
Over the same period, other AI hardware sold in the millions. The difference is that it did not ask anyone to change their life.
Ray-Ban Meta glasses passed 7 million units in 2025, more than triple the year before. The flagship, the Ray-Ban Display ($799, September 2025), adds a small in-lens screen and a wristband that reads the electrical signals of your wrist muscles to control it with a finger gesture. They work because they are glasses first: the AI is a bonus, not the only reason to wear them, and there is no leap of faith toward a device that replaces the phone.
Then there is the soberest category of all: recorders that transcribe and summarise. Plaud sells the Note (about $159), the wearable NotePin and the Note Pro, leaning on the best cloud LLMs for transcription. It is one of the few AI-hardware startups turning a profit: about $56 million in revenue in 2024 and over 2 million devices shipped by June 2026. Meanwhile the two “always-listening” rivals were absorbed by the giants: Bee by Amazon (July 2025), Limitless by Meta (December 2025). A recorder is an object you understand in a second; the AI adds value without asking you to rethink your habits. The open question is privacy, because these devices capture real speech, often continuously, and their concentration in a few large hands is anything but neutral.
The line that separates the two groups
The axis is simple: augment what exists, or replace it with something new. The Meta glasses and the Plaud recorders start from an object the user already wanted and already knew how to use, and add the AI as an advantage. Rabbit, Humane and Friend asked people to adopt a new device, with a novel way of using it (the palm projector, the “Large Action Model”, the listening pendant), whose only reason to exist was an AI that then failed to deliver, while a phone that did more was already in the user’s pocket.
The Codex Micro, in miniature, is on the right side: it is a macropad, a thing that already works, with agent status as the extra. That is why it reads as a charming curiosity rather than a bet. The Ive device, instead, is the bet: it goes back to a wholly new category, and recent history says that is the hardest road.
What we think
We like well-made objects, and we build hardware and AI together ourselves: the Codex Micro is a smart accessory, and the “augment, don’t replace” lesson is a healthy one. But for us the real point is elsewhere, and it is not on the desk.
The dial that matters is not the one adjusting an agent’s “reasoning” on a $230 pad. It is the lever over which model runs and where your data ends up. A Codex Micro is a nice control surface, but every action it launches runs on OpenAI’s cloud Codex: the same question we raised about coding CLIs applies, what actually leaves your perimeter. A key glowing green does not tell you which infrastructure your code is running on, or under what rules.
The same logic that rewards “augment what exists” hardware rewards, on the software side, those who own the operational floor instead of renting it. The layer that lasts is not a gadget: it is the model you run in-house and the control plane you govern it with. On one side open-weight models on your own hardware, as we covered for colibri, Ornith 1.0 and Inkling; on the other governance over any model, local or remote, which is what Admina does with an audit trail and ALLOW/BLOCK/REDACT policies. That is the sense of Open Intelligence, Secure Governance: you steer the fleet of agents better from a control plane you own than from a key someone else can switch off remotely, as already happened to the people who had bought a Pin.