Should I Buy Claude a Mac Mini?

Posted Jun 30, 2025 by Benjamin Anderson

I am buying Claude a Mac Mini. Or at least, I'm thinking about it.

For those of us working as software engineers, it's becoming increasingly accurate to describe language models as "co-workers" rather than tools. This is not true for everyone: people programming in arcane languages or working on hard problems that aren't in the training data still can't totally trust AIs. But if you're building a run-of-the-mill web application, and Anthropic isn't flooded with traffic and quantizing the bejeezus out of Claude Sonnet 4, it can do a substantial amount of real, useful work.

Engineers in my circles are figuring out how to adapt their development process to this new reality. Some are "rawdogging" (using Claude instances across several tabs and hoping they don't write conflicting changes to the same files), while advanced users are adopting git worktrees or specialized tools like uzi to manage their AI gofers. Soon, though, I think it will make sense to give AIs their own workstations. Here's why.

The Missing Piece for Coding Agents

The coding agents that most of us use day-to-day (Claude Code, Codex, Cursor, Windsurf) own a limited slice of the development process—they can search for and read files, make edits, and run tests. However, at least when it comes to web applications, they can't fully "close the loop", because they can't interactively test them. There are countless startups are working on AI-powered QA testing. But in my view, basic interactive testing—making sure the change you made actually did what you want—shouldn't require a separate product. Interactive testing is core to the development workflow, not an accessory. As such, integrating it into coding agents is the natural next step for all companies chasing automated software development.

Coding Agents are Computer-Use Complete

I lied earlier. Strictly speaking, coding agents like Claude Code already can do interactive testing. In fact, they can do basically anything on the computer that you can, because you, in your infinite wisdom, have decided to give them access to your shell and filesystem. Computer use is just a loop that alternates between taking a screenshot and running a shell command. AppleScript, PyAutoGUI, xdotool, etc. allow anything on your system to be done from the shell. If you like, you can add a bit more scaffolding via Model Context Protocol. But at its core, the "coding agent" abstraction already "contains" computer use. (GUI-based computer-use "contains" coding agents too, but I would not recommend this setup.)

Whether current models are good enough to manage development cycles that include interactive UI testing is another matter. But I claim the pieces are all there, and if better models are coming (and they are!), it's only a matter of time.

Your AI Needs a Computer

So, we've established that "coding agents" and "computer use" are converging on roughly the same thing—models that autonomously carry out long-range tasks on the computer. Whether they use a GUI or the terminal is beside the point. Whether they're coding or using AutoCAD or buying you shoes on Temu is honestly also beside the point. The point is that your AI will need a computer.

The most obvious objection is, why not share? After all, I and many developers are used to sharing with Claude. I work in one terminal tab, Claude works in another. I work on front-end stuff while Claude works on back-end stuff. Claude rushes out changes while I impatiently refresh the browser waiting for Claude's changes to hot-reload. This is all fine, right?

Just kidding, this is obviously a mess. And it's only going to get worse when coding agents and computer use converge, and you have to share a GUI. Sharing your work computer with an AI agent is a bad idea for the same reason that sharing it with your co-worker is a bad idea: you will get in each other's way. You can't share a mouse, it's difficult to code on one monitor while Claude tests your app in another, and you probably shouldn't run 8 development servers on 8 different ports on one Macbook Air. Claude needs his own computer!

What Kind of Computer Should You Get?

Since we all now agree that Claude can and should have little a computer as a treat, what kind of computer should that be? One option is a virtual computer inside of your computer. I'm a big fan of what the folks at c/ua are building: a lightweight, Mac-OS-native utility to run computer-use containers. Putting Claude inside a container gets him out of your way for free.

If putting a computer inside your computer is too meta or too resource-intensive, I'm told that you can also rent computers on the internet. You can just build your own desktop environment on EC2 or a serverless provider like Modal Labs. Speaking from my experience doing this, though, rolling your own desktop environment is not a particularly fun activity, unless you are one of those Linux people. I spent days customizing and extending Anthropic's reference computer-use implementation, and it was hard to get right. It makes a lot of sense to me that there are startups like Cyberdesk working on batteries-included desktop environments for AI agents.

Renting seems practical, especially if you want to spin up and spin down tons of agents. But there's something captivating about the concept of just... buying a computer. Mac Minis are very good and inexpensive. Plug one in at home or in the office, and you don't have to worry about Dockerfiles or Tint2 or NAT gateways or SSH keys or residential proxies. It has a persistent filesystem. You can back up the filesystem with Time Machine. If Claude bricks the computer, you can restore it from the backup. You can watch what it is doing by just plugging in a monitor. You can record the screen with any one of dozens of screen capture apps. You can use Tailscale (for free!) to connect to it from anywhere.

I'm just saying. It seems like... not the worst idea in the world? I don't know. I haven't tried it yet, but I want to. Maybe I just want to buy a Mac Mini and I'm rationalizing it. You tell me!

Addendum: What about Operator?

I'll cut to the chase: I do not think that a virtual browser is sufficient to do 99% of real work. One of the biggest limitations of OpenAI's Operator and other browser automation platforms like BrowserBase is that the AI is trapped inside of Chromium. This is great for security! Yay! It's also convenient to be able to take actions with Playwright instead of clicking. But the costs are severe.

Imagine if I told you to make a web application, but you had to use a McDonalds kiosk as your dev environment. That would be really hard. Real work usually involves saving files, switching out of the browser to open and edit them, using desktop applications, and so on. Even for tasks strictly limited to gathering data from the internet, how do you get the data out? Have Operator paste it line by line into a Google Sheet? That seems silly. What if I want 10,000 images of koalas downloaded and converted to WEBP? You can't do that on a kiosk.