I spent two days trying to get OpenClaw set up. The first day was pretty much hell — I started with OpenAI as the model and it kept doing the thing every developer dreads: instead of actually doing the task it would give me step-by-step instructions on how to do it myself. That completely defeats the point of OpenClaw, so after banging my head against that for a day I switched models.

I moved over to Anthropic's Claude Sonnet 4.6 and, honestly, it just works. No lecturing, no hand-holding — it does the thing instead of telling you how to do the thing. That was a relief. Once that actually cooperated the rest of the setup stopped feeling like punishment and started feeling like tinkering again.

Yes, I gave it full root on the machine. Sounds scary, I know, but that's exactly why this runs on a Raspberry Pi sitting on my desk. If it decides to go rogue the Pi can shut itself down; it's not my main machine. It's a sandboxed $50 computer that I can afford to let loose and then power-cycle if needed.

The neat part is how I use it now: via WhatsApp. I send voice notes — literally the same kind of voice note I used to draft this post — and it figures out what I want and just does it. From quick tweaks to building a full trading app, I was sending messages and it was shipping code and wiring things together. At one point it built an entire trading app for me from those messages alone. It felt a little like having a developer in the room who listens better than most humans.

I also gave it its own GitHub account. It commits, owns its own repo, and only ever touches the things I explicitly hand it. I'm even thinking of adding it as a collaborator on some personal repos so it can contribute directly — basically the setup you’d use for a hired developer, except this one lives on a Pi and doesn’t cost me a monthly fee.

There are still little quirks and things to tighten up, but getting past that awful first day made everything else feel possible. More to come — keep in touch.