Have you ever wanted to try out a new app, or open a file you weren't fully sure about, but felt nervous doing it on your main computer? What if it's broken, or worse, causes a mess you have to clean up? That's exactly the worry this new feature is designed to remove.
In computing, a sandbox is simply a separate, temporary space where you can try things out safely. Think of it like a rented test kitchen: you can cook, make a mess, and try new recipes, and when you're done, you just walk away. Nothing you do in there affects your own kitchen at home.
A "virtual machine" (VM for short) is how you create that separate space on a computer. It behaves like a whole separate computer running inside your real one, with its own operating system and its own storage. You can turn it on, use it, and throw it away without ever touching your actual machine.
We've built a way to create a full, temporary Mac desktop that runs on your own Apple Silicon Mac (any Mac with an M1, M2, M3, or M4 chip). It behaves like an ordinary Mac: a real desktop, a real Finder, and the ability to install real apps, including ones distributed as .dmg
files, the standard format for Mac installers.
The key part is that this desktop is disposable. Once you're done testing, you delete it, and everything that happened inside disappears with it. Your actual Mac stays exactly as it was.
This feature is part of SmolVM, a tool built by Celesto AI for running small, temporary virtual computers, or "sandboxes," on your own machine. SmolVM already lets people spin up disposable Linux environments, and this release extends that same idea to macOS.
You don't need to know anything about AI agents to use it, but SmolVM was built with them in mind. An AI agent that needs to test software, click through an installer, or explore a desktop can now do all of that inside a throwaway macOS sandbox instead of touching a real computer.
.dmg
installer files.This feature is used through a command-line tool, which just means typing short text commands into a "Terminal" app instead of clicking buttons. Don't worry if you've never done this before: each step below is a single line you can copy and paste.
First, a one-time setup. This downloads what's needed from Apple, which takes roughly 20 to 40 minutes and about 50GB of free space, so it's worth doing before you're in a hurry:
smolvm setup --macos
Next, create your first sandbox and give it a name:
smolvm sandbox create --os macos --name nimble-mac
Finally, open its desktop:
smolvm sandbox desktop nimble-mac
Apple's Screen Sharing app opens automatically, showing a real macOS login screen. Log in with username lume
and password lume
. This login is shared and meant only for local testing, so avoid signing into personal or sensitive accounts inside the sandbox.
If you're curious how this works, here are the details, though none of this is required to use the feature.
Each sandbox is powered by an open-source project called Lume, which uses Apple's own built-in tools for running virtual Macs. Rather than building a full copy of macOS for every sandbox, SmolVM prepares one base copy and gives each new sandbox a lightweight clone that stores only the changes you make. That's why creating a sandbox after the first one is fast, and why deleting a sandbox never affects any other sandbox or the base copy.
The connection to your sandbox's desktop only works from your own Mac and never travels out onto the internet. The security features you'd expect from a normal Mac, like Gatekeeper and System Integrity Protection, stay switched on inside the sandbox too.
A few limits apply for now: you can run at most two sandboxes at once, each one gets a fixed amount of computing power (4 CPU cores, 8GB of memory, and 80GB of storage), and there's no way yet to connect to a sandbox through the command line, only through the desktop view. These limits are expected to ease over time.
Even with those limits, it's a simple, safe way to try things on macOS, for yourself or for an AI agent, without ever putting your main computer at risk.
Learn more about SmolVM on GitHub or visit Celesto AI to see what else they're building.