← Back to blog
windowsdevboxtutorial

Windows DevBoxes That Boot in Seconds, Not Half an Hour

Javier Perez, Founder · 2026-07-01

Windows DevBoxes are the slowest thing in ClusterCode, and it is not close. Because we run Windows as a full VM — via dockur/windows on any worker with nested virtualization — a fresh DevBox has to complete an unattended Windows install before you get a desktop. That is 20 to 30 minutes of waiting. And once it finishes, you are staring at a bare Windows box with none of your tools: no Node, no Git, no editor, no AI CLIs. You install all of it, by hand, every time you launch.

If you have hardware capable of hosting Windows VMs, there is a much better way to work: prepare the machine once, snapshot it, and never wait again. We call the result a golden image, and the recipe has two levels.

Level one is preparation. You launch a Windows DevBox, wait out the one-time Windows install, and open its terminal — a real PowerShell prompt, right in the browser. From there you bootstrap Chocolatey, install Node.js, Git, VS Code, and the GitHub CLI in a single command, then add the three AI coding CLIs: Claude Code via its native installer, plus OpenAI Codex and GitHub Copilot from npm. A minute of setup, and the box has everything you would ever reach for.

Level two is where the time savings live. Instead of throwing that prepared DevBox away, you snapshot it into a reusable image directly from the dashboard — the Build an Image from a DevBox flow. For Windows, the snapshot briefly pauses the DevBox to capture a consistent disk, then bakes the entire post-install state into the image: the finished Windows install and every tool you added.

Now the math changes. Launching from that golden image does not run the Windows installer and does not install a single package — both slow steps already happened, once, and are frozen into the image. The DevBox boots from the baked disk and reaches a working desktop in seconds, with Node, Git, VS Code, and Claude Code, Codex, and Copilot already on PATH.

The snapshot even records what it found. ClusterCode detects the tools on the prepared DevBox and stamps them onto the image as capabilities, so the launch picker shows chips — Claude Code, Codex, GitHub Copilot, Node.js, Git, VS Code — and you can tell at a glance what a given golden image ships with.

The whole appeal of golden images is that the expensive work is paid once. The Windows install, the package downloads, the PATH wiring you would otherwise repeat on every launch — all of it collapses into a single snapshot step. Day to day, you just launch from the image; you only rebuild it when your toolchain actually changes.

We wrote the full recipe up — every command, plus the gotchas we hit getting the native Claude Code installer onto PATH inside a non-interactive terminal — as the Windows Golden Image guide in our docs. If you run Windows workloads on ClusterCode, it is the first thing worth setting up.

Run AI agents in isolated DevBoxes on machines you own.