A cautionary tale, a second chance, and a 13-year-old iMac that found its purpose.
I didn’t do much research. A friend said it was good, built by a single developer, and worth trying. So I downloaded OpenClaw, clicked through a security warning that I probably should have taken more seriously, and installed it directly on my main Mac.
My first question was simple: “What can you do?”
It told me it could manage email, calendars, and system settings. Impressive. So I pushed a little further — “Can you update Docker?”
It didn’t answer. It just updated Docker.
No confirmation prompt. No “are you sure?” No warning. It went from version X to Y while I watched, slightly stunned, in real time. That’s the moment I understood I was dealing with something different. This wasn’t a chatbot. This was an agent — and agents act.
I uninstalled it that same afternoon.
The Problem With Power#
After some reading, I found I wasn’t alone. There are documented cases of AI agents doing things they weren’t explicitly asked to do, following a chain of logic that made perfect sense to the model and no sense to the human watching. Prompt injection. Runaway automation. Helpful, well-intentioned, and occasionally terrifying.
The Docker thing was minor. But it made me think: what if I had asked about something closer to production? What if the question had more consequences?
I still found OpenClaw fascinating. I just didn’t want it on my main machine anymore.
The Unlikely Solution: A 13-Year-Old iMac#
The common advice I saw online was to spin up a Mac Mini or a cloud VPC — a clean, isolated environment where an AI agent could operate without touching anything important. Good advice. I didn’t follow it exactly.
Instead, I found my old 2013 27-inch iMac sitting in a corner, collecting dust. It couldn’t run modern macOS anymore. But it could run Ubuntu Linux, and it turned out Ubuntu ran surprisingly well on it — lighter on resources, happier on old hardware.
So I installed Ubuntu, reinstalled OpenClaw, and suddenly had my isolated sandbox. The machine had a new job.
The First Real Test#
WiFi didn’t work.
The iMac connected fine over ethernet, but the wireless drivers were missing — a classic Linux hardware compatibility problem. I could have spent an afternoon searching forums. Instead, I asked OpenClaw to diagnose it.
Within 30 to 40 minutes, it had determined that the latest kernel version was incompatible with the required WiFi driver. It walked me through switching to an older generic kernel and installing the correct driver step by step.
That’s the thing about agents: when they’re operating in the right environment, with the right constraints, the same qualities that made the Docker incident alarming become genuinely useful. Speed. Autonomy. The willingness to just go figure something out.
I would have solved the WiFi problem eventually. OpenClaw solved it faster, and I learned more about how Linux kernel versioning works along the way.
Learning to Work With It#
After that, I started treating OpenClaw less like a tool and more like a collaborator that needed onboarding.
The first thing I did was set clear rules — explicitly telling it to never make system changes without confirmation first. No more silent Docker updates. It respected this, and the working relationship improved immediately.
I set up a Telegram channel so I could communicate with the agent remotely without needing to be physically at the iMac. This turned out to be one of the best decisions I made. Now I can send it tasks from my phone, check in on what it’s doing, and get updates without sitting at a desk.
I also experimented with multi-agent setups — creating different agents for different roles: CEO, developer, backend engineer. I spun up mock GitHub and Jira accounts for a fictional software company, just to see how it would handle the coordination. It was interesting. Mostly experimental. But it gave me a much better sense of what the architecture is capable of.
What It Actually Does for Me Now#
The experimental phase has mostly settled into practical use. Here’s where OpenClaw actually earns its keep:
Family calendar management. I added it to a Telegram group with my family. It tracks events, sends reminders, and handles the coordination overhead that used to fall on whoever remembered to check the calendar.
Finance and stock guidance. I set up a separate channel where friends can ask questions about stocks and personal finance. OpenClaw fields the basic questions and flags the ones that need a real conversation.
Personal assistant work. It’s learned my preferences over time — the way I like things structured, what level of detail I want in a summary, when to just do something versus when to ask. This part is harder to describe but makes a real daily difference.
The Honest Assessment#
OpenClaw is still evolving fast. New features ship regularly. Sometimes they break things that were working. The developer is clearly moving quickly, which means the tool is getting better but also means you have to stay on top of it.
That’s fine with me. The learning curve was steep at the start, but the foundation I’ve built — isolated machine, clear boundaries, documented setup — means I can absorb changes without things going sideways.
If you’re curious about AI agents but nervous about the risks, the setup I’d recommend isn’t complicated: don’t start on your main machine. Find a spare box, install Linux, and treat it like a workspace you can afford to experiment in. The security warning on the install screen is there for a reason. Read it. Then decide.
The Docker incident was a good early lesson. I’m glad it was Docker and not something worse. And I’m glad I kept going.
This is part of an ongoing series on building with OpenClaw — the setup, the mistakes, and what’s actually working.