Local first, cloud ready
Every agent needs a computer. This is its.
Make things: the software you build and test, the content you produce, the routines that keep a business running. Work that happens in real apps and on real websites, done with an agent in an isolated Linux desktop, on your real files. Watch it, take the wheel, or design the process once and get back proof.
Apple silicon Macs · one email when the early build is ready
Use cases
Describe the result. The agent lays out the process and runs it in copies — with proof.
Some jobs come back every week: check a build, retest a bug, compare two versions. Lay one out once, and let the board run it.
01 You drivetoday
A Linux desktop that opens like an app
Chromium, Blender, VS Code — anything in the catalog opens in its own copy: a NixOS microVM with its own compositor, rendering on the Apple GPU. Your project mounts read-write. Tile a few. No agent involved. Next: snapshot a copy at any point and come back to exactly that state.
Setup — none
02 You delegatetoday
Hand a task to an agent. Keep the window open.
Type the goal in the chat. The agent works inside the copy over MCP, and you watch it in a real window. Pause after the current step, steer with a message, or click in and take the wheel. Claude and ChatGPT work out of the box, or pick your own model — MLX or GGUF, running on your Mac.
Setup — a chat
03 You designnext
Describe the result. The agent lays out the process.
The board above. Today, a process is a markdown plan you @-mention in a chat, and it already runs; the canvas, the trace, and the proofs wired to each step are what we're building next.
Setup — an afternoon
Where it runs
A macOS app on Apple silicon. Each copy is a NixOS microVM with its own compositor, rendering on the Apple GPU.
What is inside a copy
Apps arrive from a shared Nix store, identical in every copy. Snapshots freeze a copy at a checkpoint and start N copies from it.coming next
How agents drive it
Over MCP: spawn · screenshot · click · type · exec · steer. Bring your own agent — its.computer is the computer, not the brain.
Models
Claude and ChatGPT out of the box, or your own MLX / GGUF model running locally.
An agent shouldn't just have tools. It should have a computer.
It's a computer.
Give your agent one. Early builds, warts and all.
Apple silicon Macs · one email when the early build is ready
Where this goes · local-first, cloud-ready
Infrastructure for agents to operate the same apps and computers humans use
Give an agent a computer
Real apps, real files, on your Mac. You watch, steer, take the wheel. Nothing leaves your disk unless you send it.
Give agents reproducible computers
The environment from a Nix flake, snapshots to fork from, the board to design the process once.
Run many — and check the work
N copies, N agents, N scenarios. The same copies validate the result against the design, replay it on a PR, and flag what doesn't hold — you decide.
Cloud ready
A process moves to the cloud with an address of its own — alice.its.computer — and an inbox. Reach it from Claude, ChatGPT or mail; it speaks MCP and email. Nothing you design is published by default.