#The macOS sandbox I actually want

I tweeted this yesterday and it got longer than a tweet.

I still think we need better Mac OS vm / sandbox experience which are equal to how Linux vms are as of now.

Different ways different team solving it. Mac satellite, own runners from ampcode. Other providers coming in like blacksmith, use computer.

Tuist seems to be better…

— Rizwan (@rizzu26) November 7, 2026

On Linux I spin up a VM, let the agent work inside it, and delete it when done. Nobody thinks about it anymore.

macOS is not there. I wrote about the license limits and prices in September: two VMs per Mac, and a 24 hour minimum when someone else hosts the Mac. Six weeks later both rules are still in Apple’s license. The difference now is that I need one.

Who is building it

• Amp: macOS Satellites. The thread stays on a Linux orb. When a task needs macOS, Amp starts a Mac with Xcode and the iOS simulators in about 30 seconds and copies changed files over before each command. The Mac app can also run threads on this Mac, which is close to what I do today with a mini at home.
• Blacksmith: agent sandboxes in beta, full VMs that match their CI environment. Still a waitlist, same as in September.
• Daytona: macOS sandboxes, and where use.computer went (the old links redirect there). Reserve an M4 Pro Mac for 24 hours and run up to two sandboxes on it. Xcode and the simulator come in the image, and the agent drives apps through screenshots, mouse and keyboard.
• Tuist: runners went public this week. Fresh macOS VMs, Xcode compilation cache, and an MCP server so the agent can read job logs. CI today, but the closest thing to a sandbox from someone who only cares about Xcode.

What I want

• Restore, not boot. Come back from a saved state in under a second instead of a cold boot every time.
• One Mac per task, thrown away after. Each agent gets a clean VM, and nothing leaks into the next one.
• Builds that start warm. Xcode is already there, and the image has built the project once, so the first build is not a cold one.
• The agent runs inside the Mac, next to the simulator. Not xcodebuild over SSH from a Linux box.
• Agents talk to it over MCP. Lease a Mac, run a command, copy files, release. Claude Code or Amp picks it up without me writing glue.
• It runs on Macs I already own. Point it at a mini at home and pay nothing per hour, instead of renting by the day.

And the wishlist on top: pause a VM with Xcode open and resume it later, or fork a warm one when I want to try a second branch.

Nobody has all of this yet.

The bill if you leave it on

Daytona charges $60.08 for a 24 hour Mac. Leave it on for a month:

$60.08 × 30 = $1,802

That Mac runs two sandboxes, so about $900 each. A new M6 Mac mini starts at $899.

Renting a bare metal mini is around $100 a month, sometimes more. After nine months you have paid for one.

A hosted Mac is easier to swap when a new chip comes out, and a dead disk is not my problem. A mini at home is cheaper to leave running, but the upgrades are on me.

Why I care now

My laptop is a 2021 M1 Pro with 32 GB and 1 TB. As I write this:

• 80 GB free, the disk is 92% full
• 12 simulators taking 19 GB
• 1.9 GB of swap in use
• DerivedData at 0 bytes, because I cleared it again

Claude Code and Amp run on it all day. Every feature gets a worktree and its own simulator. RAM runs out first, then disk. I clean storage almost every week and still cannot keep a few feature simulators running at the same time.

I do not want another mini on my desk, and I do not want a cloud bill the size of a new Mac. So for now I will keep the setup from September: rent a cloud Mac for short jobs, and use the mini at home for everything else. Next I want to try Amp Satellites on a real feature and see if it takes that load off this laptop.