Computers are still chosen from a list.
Companies issue computers by the model number, and each person’s work is fitted to whatever is on the list. People who build their own go one part at a time, and a good price on one part decides more of the build than anyone meant. Either way, the computer is decided before anyone writes down how it will be used.
Meanwhile more of the work runs as agents: long jobs, large models, data that should stay inside the building. A machine chosen for an average user runs hot, runs down and sends the heavy work to someone else’s data center.
A good deal should make sense for the whole build.
We believe a computer should be designed from its brief, and the brief includes how it is used: where it belongs, what its owner carries, what the work must keep private. Price and performance matter. They are not the only things worth designing around.
The aim is a design you understand, not a list you are expected to trust. Every choice carries a written reason, and every rejected option the reason it lost. The software must also be free to say buy nothing, when the machine you own, a smaller model or a hosted service is the better answer.
An agent to propose, physics to check, a person to decide.
Designing a computer as one system has always taken a team: sourcing, thermal, mechanical, industrial. That cost is why computers come in a few fixed models, and why a design for one team or one person rarely pays for itself.
Agents change the cost of the first pass. The agent we are building will work with a parts catalog, a physics solver and parametric CAD. It will propose changes and save nothing on its own. It will show what it ruled out and why, and where each specification came from. It will estimate throughput, power and heat before anything is built, and hold back fabrication CAD until the design is ready. At each decision that matters, it will stop and ask.
Why now.
Three things changed together. Capable open-weight models now fit on a machine you can carry. Processors with 128 GB of memory shared between CPU and graphics now ship in mini PCs. And agents can now use engineering tools, not just describe them. The first two make a personal machine worth designing around its work. The third makes the design affordable.
One line we hold: a machine designed with Outfra can keep its data where it runs. The design agent itself will run in the cloud.
The first machine is the briefcase in the drawing.
We are starting with work that has a reason to stay local: a boundary on the data, a site without a connection, a bill that grows with every token. The first design is a briefcase workstation for running agents, with its own cooling, battery and keyboard, that plugs into the monitors at a desk without being opened. It is the machine a company hands a person when a rack cannot go with them.
Our solver already predicts its throughput, power and heat. Building the first prototype is what calibrates it.
Then every computer with a brief.
The same process leads elsewhere for a machine that lives on a bench, rides in a vehicle or sits in a room with no connection. For a company, that means a fleet specified by the work each role does, not by model number. For a person, a build that makes sense as a whole before the first part is bought. Either way, the computer starts with how you want to use it.
Have a computer in mind?
We are looking for early users with a specific workload and a budget: teams that issue machines to people, and people who build their own.
Access is granted by invitation.