The Fountain Institute is an independent online school that teaches advanced UX & product skills.
I Bought a Mac Mini to Try OpenClaw, the Most Hyped AI Tool of 2026
I Bought a Mac Mini to Try OpenClaw, the Most Hyped AI Tool of 2026by Jeff Humble Dear Reader, You've probably heard of OpenClaw π¦ by now. 145,000 GitHub stars. Headlines everywhere. "The AI that actually does things." This tool is the O.G. dream of AI...automation, not slop. This was the missing piece to my automation system. I had to try it. So I bought an entry-level, 2024 M4 Mac Mini for β¬590 (on sale in Germany, but they're reportedly selling out in the U.S.) and spent two days trying to get OpenClaw running. Here's what actually happened. Why I Did ThisI spend too much time on little piddly things that an AI could probably do better than me. For example, I probably spend an hour every week doing my invoices, and the time saved on just that would pay for the machine... And hey, I needed a new computer for running live classes, anyway. So the math made sense to me. And I loved the idea of running open-source models locally and replacing Zapier for nothing but the cost of the cheapest Mac and some extra electricity. But really, I also wanted to see what all the fuss was about... Setting Up OpenClaw (Using AI to Set Up AI)Here's where it gets weird. I tried the normal install, but it's pretty technical and didn't work on the first try. So, I used Claude Code to prompt Terminal to diagnose and solve the problem. Then I cross-checked those commands with ChatGPT's free version to make sure I wasn't accidentally giving the AI access to things it shouldn't have. One term kept coming up: the principle of least privilege. Basically, only give the AI the minimum access it needs to do its job. So I set up OpenClaw as a guest user on the Mac Mini, made myself the admin, and created a dedicated Gmail account just for it. The AI got its own sandbox, and I kept the keys in case the little π¦ ruined my Mac Mini sandcastle. This back-and-forth (between Terminal, OpenClaw's gateway, Claude Code, and ChatGPT) went on for hours. I was using AI to install AI while asking a different AI if the first AI could be trusted. Peak 2026... The Reality CheckI wanted OpenClaw to run entirely on local models, meaning everything stays on my machine: no API costs, totally free after the hardware. Claude recommended several models that should work on the M4 Mac Mini. They were all about twice as slow as predicted and basically unusable... I'm still trying to figure out why. After hours of trying to optimize, I gave up on local-only and switched to Haiku, Claude's cheaper API model. That's when I burned through $3.14 in an hour, mostly just troubleshooting and testing (not even doing real work). OpenClaw packages your conversation history into every query, so the token costs add up faster than you'd expect. For context: my Claude Pro subscription is β¬20/month for 10x the speed and capability. In one hour of poking around OpenClaw, I spent $3.14 doing essentially nothing productive. The economics of AI agents are real, and most of us have no idea what these tools actually cost per query when you're paying at the API level. I might go back and try to optimize the local models, or switch to MiniMax for about $10/month, like others in the community have done. I really want to believe that the dream of "free, private, local AI" is possible in 2026, but on an entry-level machine, it ain't there yet. Then My Son Came Home From DaycareAfter hammering my head against the wall all day, I was forced to walk away. And honestly, that's when it started working for me. The WhatsApp integration was as simple as scanning a QR code generated by Terminal. Suddenly, I was messaging myself to chat with my AI. No Terminal. No gateway. Just a WhatsApp thread. At one point it somehow entered a chat with a friend, but I told it to stop doing that. Let's see if he listens. In the WhatsApp UI, OpenClaw felt totally different: like texting a Clippy burrowed deep into your machine. Except this one has persistent memory, knows what I've been working on, and can actually do things on my computer when I ask it to...theoretically ππΌ That shift, from fighting with Terminal to casually texting my AI from the couch, was the moment OpenClaw clicked for me. What I Want to Try Next with OpenClawI haven't gotten into the real actions yet. The agent skills (the things that make OpenClaw more than a chatbot) are still ahead of me. But here's what I want to test, and these are things I think any product designer could use:
These are the use cases I'll be exploring in Part 2. I want to find out if OpenClaw can actually do this stuff reliably, or if the hype outpaces the reality. All in the Name of EfficiencyTotal cost of this experiment: β¬590 for the Mac Mini, β¬20/month for Claude Pro, $3.14 in API tokens, and two full days of my life. Plus the free versions of ChatGPT and Google's AI search results to keep me safe along the way. Was it worth it? OpenClaw is the custom motorcycle of AI assistants: hard to keep running, but can be a work of art if you have the time to tinker. Claude Code is the electric scooter you rent from your phone: anyone can access it, but it's not serious transportation. Both will certainly get you to some interesting places. At one point I had Claude Code, ChatGPT, and OpenClaw all open at once, each one checking the other's work. Nobody was in charge. It was incredibly efficient. And I'm not sure if that's a compliment. More in Part 2... Cinematic slop: Just in case you thought it was only designers who should worry about their jobs, there's a new video model that is making Hollywood quiver. It's called Seedance 2.0, and it's shockingly good. Videos of Tom Cruise fighting Brad Pitt helped fuel the internet storm.
β
It's also pretty good at more subtle movement and emotions. Notice the allllmost completely normal hands on this guy.
β
The model comes from Bytedance (company behind TikTok), so a social media integration is probably in the works...unfortunately. from the Fountain Institute YouTube channelβ
Like hearing about practical AI stuff? Or want to hear less about AI? Hit "reply" and let me know what you think!
|
The Fountain Institute
The Fountain Institute is an independent online school that teaches advanced UX & product skills.
Why Decisions Feel So Hard Right Nows By Hannah Baker Dear Reader, Over the last few months, Iβve been talking with design and product leaders across very different organizations, large companies, smaller teams, fast-moving environments, and slower ones. And I keep hearing the same thing. Their teams are being asked to make decisions faster than ever, and yet, deciding feels heavier than it used to. Not slower, exactly. Just harder. At first, people often explain this in familiar ways: too...
When Speed Stops Being the Bottleneck by Jeff Humble Dear Reader, Quick question: What happens when the thing that used to take 12 weeks now takes 4 days? I've been watching this play out across the industry, and it's wild. Lots of companies aren't sharing their new speeds, but a few are: Code and Theory (an agency that works with Microsoft and Amazon) is building dashboards in 40 minutes that used to take a week. They report cutting time-to-prototype by 75%. Coinbase reports a 2-5x increase...
Missed a few newsletters last year? Start here. By Hannah Baker Dear Reader, Before we jump into new ideas for the year ahead, we wanted to pause for a moment. If your inbox was anything like ours last year, thereβs a good chance you missed a few newsletters. So instead of sending another new idea right away, we put together a curated catch-up, a handful of pieces from 2025 that capture the questions we kept returning to. If you only read a few things from us last year, these are a good place...