Running an AI Agent on a Raspberry Pi Just Because I Could

I had a Raspberry Pi 3 sitting around doing absolutely nothing, so naturally the first idea was to try running one of the new AI computer-use style agents on it. Not because it made sense. Just because I wanted to see if it would work.

The setup was intentionally minimal. Fresh Raspberry Pi OS install, a couple system packages for screenshots and mouse control, Python, and the OpenAI client. The Pi itself is not running any AI models. It just acts as a controller that sends requests to the API and executes whatever actions come back.

The entire point of the experiment was isolation. There was zero chance I was installing experimental automation software on anything I actually care about. The Pi lives on its own network and if it melts down, great. That is literally the job.

After wiring everything up I ran the first test script and immediately hit a rate limit error. Which, hilariously, was actually the success signal. The script executed correctly, reached the API, and failed only because billing was not enabled. That meant the controller side of the setup was working perfectly.

From there the rest is just plumbing. The Pi can take screenshots, move the mouse, and click things. The AI just decides what to do next. Slow? Absolutely. Useful? Debatable. Funny that it works at all on a tiny $35 computer from years ago? Definitely.

Sometimes the best projects are the ones where the goal is simply “I wonder if this will run on that.”