A Mac Mini for OpenClaw, and Why I'm Already Drifting to Claude

There's a new Mac mini on my desk, and it was bought for exactly one job: to be the always-on brain of an AI agent I could talk to from anywhere. A week later, Anthropic shipped two features that make me wonder if I timed that purchase about as well as a snowplow in July.

The plan, when I ordered the thing, was simple. I'd been running OpenClaw, Peter Steinberger's open-source gateway that bridges Slack, Telegram, Discord, and a pile of other chat surfaces to Claude Code, on my laptop. Which works, right up until the laptop closes, or sleeps, or travels somewhere I'd rather leave it alone. The whole idea of an always-on agent is that it's always on. A machine that goes where I go is a machine that's always kind of off.

A Mac mini solves that in the dumbest, most satisfying way: a small silent box under the desk that doesn't close, doesn't sleep, doesn't come with me to coffee shops. I bought it specifically for rocky, my OpenClaw agent. Rocky is not a person. He's a separate user on the machine with his own home directory, his own keys, his own sandboxed little life. The whole point was that he could be always on without having access to my personal files, my emails, my everything. He got a dedicated space, and I built him some genuinely useful tools. He served well.

I was in St. Thomas when it hit me what this setup could actually do. Power was out at the rental, I had barely enough cell service to load a Telegram message, and my laptop was back on the mainland. None of that mattered. I thumbed Rocky a quick task from my phone, just a few words, and he picked it up back home and ran with it. By the time I had a drink in my hand he'd sent the results back. All I needed was enough signal to get a sentence out, and a computer 2,000 miles away did the rest.

The day-to-day reality was a little bumpier though. OpenClaw is fast-moving and occasionally feral. Every few days something would break, a plugin would go sideways, or a message I fired off from Telegram at 8 a.m. would quietly fail to route until I noticed at noon. The community around it is shipping at a genuinely impressive pace, but "impressive pace" and "rock-solid infra" aren't usually the same sentence. You learn to treat Rocky like a houseplant: interesting, rewarding, needy.

And then, over the last week, Anthropic did the thing where they just… ship the managed version of what I was building with duct tape.

On April 9, Cowork went generally available, a desktop-first Claude Code experience for people who don't live in a terminal, with scheduled and on-demand tasks baked in and a proper spot to manage skills, plugins, and connectors. Then yesterday, April 14, they dropped Routines: saved Claude Code configs that run on Anthropic's own infrastructure on a schedule, from an API call, or triggered by a GitHub event. No laptop required. No Mac mini required. No houseplant.

When I squint at Routines, it is uncannily close to the thing I bought the Mac mini for in the first place. "Dynamic cron jobs for agents" is more or less how I'd describe what I wanted OpenClaw to be for me. A thing that wakes up, looks at the world, does the work, goes back to sleep, and doesn't ask me to be anywhere in particular while it's doing it. I'd been assembling that out of open-source parts on hardware I own. Anthropic is now renting me the same shape of thing, managed, with better uptime than my electrical outlet.

The center of gravity is shifting, one small task at a time.

So the Mac mini isn't going anywhere, but Rocky is going on vacation. Now that Anthropic has shipped the managed version of what I was hand-wiring, I'm testing out the same always-on setup on my own user: alex, logged in and running on the same little box, but leaning on Claude's stack instead of OpenClaw's. Same hardware, different tenant. The Mac mini stays; the occupant changes.

On the alex side I've been leaning into Dispatch, Anthropic's new thing that lets me fling a task at Claude from my phone and pick up the results on the desktop when I'm back at it. When I want something reliable, a scheduled repo check, a recurring digest, a task I don't want to babysit, I'm reaching for Dispatch, Cowork, and Routines instead of SSHing into Rocky's account to restart a gateway at 11 p.m. The menial stuff that Rocky used to fumble through on a good day just… works. There's something philosophically useful about having the self-hosted path exist, and honestly, tinkering with OpenClaw was fun in the way that debugging models was fun back in my early data science days. You learn a ton. You feel like you own the thing. But at some point you want it to just work so you can focus on what you're actually building.

It's starting to feel like the thing Rocky taught me wasn't how to run an always-on agent. It was what I actually wanted an always-on agent to feel like. Now that Anthropic is starting to fit the bill, Rocky's user account is gradually going quiet, a monument to the couple of months where "always on" meant "I built it myself." He earned his vacation.

And the alchemist keeps turning lead into gold. Just with slightly fewer bash scripts this time.