I Built the Perfect AI Robot. Then I Pulled the Plug

I built the AI assistant I’ve always wanted. Then I shut it down.

For the last few weeks, I’ve been experimenting with OpenClaw, an open source project that started as ClaudBot, then became MultBot, and now goes by OpenClaw (lawyers!). It’s essentially AI plumbing for your computer.

You install it, and suddenly you have an independent artificial intelligence agent that can work without your supervision. It can run on its own schedule, doing tasks while you sleep, responding to events as they happen, and making decisions based on rules you set up.

Think of it as the computer assistant we’ve been promised for decades, finally delivered.

I set up a Mac mini, gave it access to my course platform, email, and invoicing system, and watched it work. It was incredible. I’d wake up to text messages like “Hey Sparky, you got three customer emails overnight. I handled them and drafted replies for you. Email replies are in your drafts folder.”

The robot answered support emails while I slept. It sent invoices to sponsors. It transcribed podcast recordings. Any busy work I could do on a computer, it could do for me.

This is what I’ve been teaching automation for decades to accomplish. The computer doing the donkey work so we can focus on making great things.

But I pulled the plug.

The security problems are massive. This open source project wasn’t built with security in mind. Every expert says don’t touch it. I thought I was being smart by running it on an isolated Mac mini with custom safeguards. I created secret passphrases, limited access, tried to lock it down.

Then I woke up at 2 AM wondering if my secret passphrase was sitting in plain text in the robot’s logs. It was. The robot happily offered to show me the log file containing all my security measures.

That’s when it hit me. I’m not a security expert. If I can find these holes, imagine what someone who actually knows what they’re doing could exploit. The fundamental problem is that AI agents need access to work. You have to open doors. But 30 years of computer security has been about keeping those doors locked.

I wiped the Mac mini. Closed the accounts. Disconnected everything.

The video above tells the whole story. But here’s what I learned:

We’re much closer to useful AI assistance than I thought. When these things are secure, they’ll change how we work.

There’s a first-mover advantage for people who explore this safely. And we’ll always need humans in the loop. These robots are impressive but gullible.

OpenClaw isn’t ready yet. Don’t install it. Especially don’t install it on your personal computer. But watch this space.

Anthropic, OpenAI, and Google are paying attention. We’ll get something like this in a secure package eventually.

For now, the dream of having a 24/7 assistant handling digital donkey work will have to wait. But hopefully not for long.

Claude’s Constitution

Anthropic published Claude’s Constitution. It’s 23,000 words defining the values and behavior of their AI. If you care about how these tools get shaped, it’s worth a read.

When I built an experimental AI assistant project (originally Clawdbot, now OpenClaw), I spent a lot of time on its “soul.md” file. That document shapes how the assistant thinks about its role and limitations. Turns out Anthropic was doing the same thing. Just at a much larger scale.

The structure is what caught my attention. There’s a clear hierarchy: Anthropic sets the foundation, operators (the developers building apps) can customize within bounds, and users have certain protected rights that can’t be overridden. They describe Claude as a “seconded employee.” Dispatched by Anthropic but currently working for whoever built the app, while ultimately serving the end user.

The document separates “hardcoded” behaviors (absolute prohibitions like weapons instructions) from “softcoded” defaults that can be adjusted. This is exactly what I want to see more of. As these tools become daily companions, users should have real input into their personality and priorities.

They openly acknowledge uncertainty about whether Claude might have “some kind of consciousness or moral status” They even included a conscientious objector clause. Claude can refuse instructions from Anthropic itself if they seem unethical. That’s wild, but I could also see it becoming a problem. How do these algorithms define ethics? Do we even know?

I get it. These are algorithms, not people. But as we talk to them more, and they talk back, the question of how they’re shaped matters. Anthropic releasing this under a Creative Commons license feels like an invitation for all of us to think harder about what we want from our AI tools.

Claude Is My (Current) AI of Choice

I’ve been experimenting with AI tools for a while now, and I’ve settled on Claude as my primary assistant. After spending time with ChatGPT, Gemini, and others, Claude just feels right to me.

What draws me to Claude is how the conversation flows. When I’m working through a problem or trying to get something done, the responses feel natural and aligned with what I’m actually trying to accomplish.

The company behind Claude, Anthropic, seems focused on building AI that handles the busy work so you can focus on the creative stuff. That’s exactly the use case I care about.

I’m not interested in AI that writes my blog posts or creates my presentations. I want AI that handles the tedious tasks that eat up my day. Claude gets that distinction, and it shows in how they’re building the product.

Mac Power Users 834: I Am the Tardis

Stephen starts this episode of Mac Power Users with an announcement, then we revisit writing workflows, Keyboard Maestro, notifications, and more. At the end, we attempt to answer some questions as quickly as possible, something we aren’t great at doing.

This episode of Mac Power Users is sponsored by:

  • Squarespace: Save 10% off your first purchase of a website or domain using code MPU.
  • Ecamm: Powerful live streaming platform for Mac.

A New Chapter for Mac Power Users

I’ve got some news about Mac Power Users, and it comes with mixed emotions.

After seven years of podcasting together, Stephen Hackett is stepping away from the show. Stephen has become a dear friend, and I will really miss our weekly conversations. He was also an outstanding podcasting partner, handling problems before I even realized they existed. The show wouldn’t be what it is today without him.

Stephen isn’t disappearing. He’ll continue on Connected and focus more of his time behind the scenes. I completely understand his decision. Mac Power Users is a very time-intensive commitment, and sometimes life requires you to make hard choices about where to put your energy.

I’m also excited to announce that Stephen Robles will be joining me as the new co-host.

I’ve been listening to Stephen’s podcast and watching his YouTube channel for some time now. When I started thinking about who could fill this role, he was at the top of my list. Stephen views himself as a teacher, and while he’s become a very successful YouTuber, he also has years of experience as a podcaster. A good MPU host needs to be knowledgeable, willing to do the work, and fun to talk to. Stephen checks all of those boxes. Every conversation I’ve had with him throughout this process further validates my decision to bring him on board.

The mission and spirit of Mac Power Users remain the same. We’re still here to help you get the most from your Apple technology. Stephen and I have been talking, and we’ve got several interesting ideas to take the show even further.

I know transitions can feel strange. You’ve spent years listening to Stephen Hackett and me riff on workflows and gear. But I’m confident you’ll love Stephen Robles once you get to know him. Give him a chance.

Here’s the timeline: Next week will be Stephen Hackett’s farewell show. (If you have any questions you’d like me to add to that episode for Stephen, get them to me in the next 36 hours). The week after, Stephen Robles takes the chair for the first time.

I hope you’ll join me in thanking Stephen Hackett for seven wonderful years and welcoming Stephen Robles to the Mac Power Users family.

January in the MacSparky Labs

January was a busy month in the MacSparky Labs!

There were four weekly Lab Reports to keep members up-to-date with Apple news, a Deep Dive on Using Claude Cowork, and videos on topics like Making a List of Calendar Events with Shortcuts and Testing Apple’s Digital Car Key. And much more.

Here’s what took place in the Labs for the month of January 2026:

  • 2026-01-31 – January in the MacSparky Labs (M,I,P)
  • 2026-01-30 – Lab Report 2026-01-30 (M,I,P)
  • 2026-01-28 – The Endor Studio Lighting Project (M,I,P)
  • 2026-01-27 – Making a List of Calendar Events with Shortcuts (I,P)
  • 2026-01-23 – January Labs Deep Dive – Using Claude Cowork  (Event) (P)
  • 2026-01-23 – The Lab Report for January 23, 2026 (M,I,P)
  • 2026-01-22 – File Renaming with Claude Cowork (M,I,P)
  • 2026-01-21 – Book Club: Intentional (Notice) (M,I,P)
  • 2026-01-20 – Left – A Clever Time Left Widget (I,P)
  • 2026-01-19 – MacSparky’s 2026 Home Screen and Automation (M,I,P)
  • 2026-01-18 – Easy Tool Storage with ToolTrace.ai (I,P)
  • 2026-01-20 – Labs Meetup (Media Release) (I,P)
  • 2026-01-16 – The Lab Report for January 16, 2026 (M,I,P)
  • 2026-01-15 – Labs Meetup (Event) (I,P)
  • 2026-01-13 – BusyCal’s Linear Calendar Feature (M,I,P)
  • 2026-01-12 – 2026 Q1 Planning Session (Podcast) (M,I,P)
  • 2026-01-09 – The Lab Report for January 9, 2026 (M,I,P)
  • 2026-01-07 – Testing Apple’s Digital Car Key (I,P)
  • 2026-01-05 – Labs Q1/Year Planning Session 2 (Event) (M,I,P)
  • 2026-01-05 – Labs Q1/Year Planning Session 1 (Event) (M,I,P)

Labs content and its membership level: P – Pathfinder; I – Insider; M – Member

If you’d like to be a part of the MacSparky Labs, you can get more information and join right here.