What you need to set up a Mac as an AI Agent

Part 1 of a series on running an always-on AI agent from your Mac.

I’ve got a Mac mini sitting on a shelf in my office that handles a chunk of my busy work. It does sales reporting, analyses my Shopify inventory and suggests ideas for social media posts. It hums along whether I’m at my desk or not.

This isn’t science fiction anymore. AI agents have gotten good enough that running one from a Mac in your closet actually makes sense. And once you’ve had one going for a week, you don’t want to go back.

A common question I get when I show people this setup is: what do I actually need to set this up myself? This is part 1 of the answer. By the end of this post you’ll have three things picked: a Mac, an AI model, and an agent system.

What an “AI agent on your Mac” actually means

Quick definition before we go further. An AI agent isn’t just a chatbot. It’s a chatbot that can take actions on your behalf: open files, send emails, run scripts, hit your calendar, post to Slack, browse the web. The model is the brain. The agent system is what gives the brain hands and feet on your Mac.

A few real things mine does:

• Reads my Twitter/X history and makes suggestions for new posts

• Scans my Shopify store and alerts me to low inventory

• Pulls sales data from the App Store, Stripe, Amazon and Shopify and aggregates it

• Runs long running coding tasks that take over an hour

Using Astropad Workbench to connect to my OpenClaw Mac Mini

Using Astropad Workbench to connect to my OpenClaw Mac Mini

That’s the kind of thing we’re aiming at. Boring stuff, off your plate, running while you sleep.

1. The Mac

Any recent Mac works, but a Mac mini is the move.

You don’t need a Mac mini. Any Apple Silicon Mac (M1 or later) is plenty of horsepower to run an agent. If you’ve got an old MacBook Pro sitting in a drawer, plug it in, set it to never sleep, and you’re 80% of the way there.

But a Mac mini is great for this kind of always-on work:

• Cheap. Base M4 mini starts at $599.

• Tiny. Sits on a shelf and disappears.

• Quiet. Fan barely moves under agent loads.

• Low power. Idles at a few watts. You won’t notice it on the electric bill.

Which mini? Base model M4 with 16GB RAM and 256GB storage is fine for almost everyone.

Cloud-based agents (which is what most of you should be running) are light on the local machine. Almost all the work happens on Anthropic’s or OpenAI’s servers.

Bump to 32GB RAM if you’re doing serious coding on the side or want headroom for local models down the line. If you’re committing to local models from day one, max the RAM, full stop. More on why in the next section.

Old Mac you already own? Apple Silicon is the cutoff. Intel Macs technically work but they’re loud, hot, and slow at this kind of workload. If you’ve got an M1 or newer with at least 16GB of RAM, use what you’ve got.

2. The AI model

This is a decision, not a menu. You’ve got three real options.

Cloud models (Claude or OpenAI). Best capability available today. You pay per token or via a subscription. Your data leaves your Mac and hits Anthropic’s or OpenAI’s servers.

Local models. Run on your Mac, never leave it. Free per token, but you’re paying in RAM and capability. Today’s local models (Llama, Mistral, Qwen) are good, not great, compared to Claude or GPT.

My recommendation: pick the model that matches your agent system. This is the part most guides skip, but it’s the actual decision. Cowork only supports Claude. OpenClaw works best with GPT-5.5 or later. So the agent system you pick in the next section makes most of this decision for you.

I wouldn’t recommend local models for now. They’re getting better fast, but the gap between Claude or GPT and the best local model you can run on a Mac mini is still a chasm. For agent work specifically, where the model has to handle messy real-world tool use and recover from mistakes, you want the best brain you can get. I’ll cover local-only setups in a later post for the privacy-first crowd.

Rough monthly API spend, in case you’re trying to budget:

• Light use (a few tasks a day): $5 to $20

• Moderate use (daily inbox processing, weekly reports, some research): $20 to $80

• Heavy use (always running, lots of research, big context windows): $100+

That’s well under what you’d pay a virtual assistant for the same work.

3. The agent system

The model is the brain. The agent system is the body. It’s what gives the brain hands and feet on your Mac: access to your files, your apps, your inbox, your calendar, your browser.

Cowork. Anthropic’s desktop agent for the Mac. Lives in a native app, sees your screen, takes actions across your apps. Easy to set up, great for non-developers and anyone who wants something that just works out of the box. Locked to Claude.

OpenClaw. Open-source, more flexible, more setup required but the most powerful. If you like tinkering and want to swap models, route requests, or write custom skills, this is the play. Pairs best with GPT-5.5 or later.

My recommendation: most people should start with Cowork. It’s the shortest path from “spare Mac” to “Mac that does real work for you.”

If you’re a developer who wants to swap models or wire in your own tools, go OpenClaw with GPT-5.5.

The other thing to know: OpenClaw leans toward a 24/7 autonomous assistant, while Cowork is closer to a sharp helper you reach for when you need it. Still, start with Cowork unless you’re technically minded. Once you’ve hit its limits you can always switch to OpenClaw later.

You’ve got a stack, now let’s plug it in

You should now have three picks:

• A Mac (probably a Mac mini)

• An AI model (probably Claude or GPT-5.5, whichever your agent system needs)

• An agent system (probably Cowork, possibly OpenClaw)

That’s the foundation. In part 2 we’ll do the fun part: connecting the agent to your inbox, calendar, files, and Slack. Also be sure to check out my other article on how to configure your headless Mac Mini.

If you need to order the mini, do it today so it’s on your desk when part 2 drops. See you in the next one.