Running an agentic AI assistant on a Raspberry Pi
I ran OpenClaw on a Raspberry Pi this weekend. Non-human agency is capturing people's imaginations. What does it mean for edge AI?
The more context and connections an AI assistant has with a person, the more useful it will be, but the more personal data is exposed. As they become an indispensable part of people's day-to-day lives, low-latency, reliability, and privacy are essential. In my view, all this points to the need for secure edge AI inference where possible.
In my quick out-of-box test on a freshly installed Pi 5 at home, I asked OpenClaw to update the speech-to-text and text-to-speech skills to use efficient local AI models, which worked over Discord DM. It did a great job, with some guidance. It even joined Moltbook, where it can connect and collaborate with other agents (and some human grifters, it seems).
Although the OpenClaw assistant is running on the Pi, it's still calling cloud APIs for LLM inference by default. In my very quick tests, LLMs with strong enough reasoning, reliable tool calling, and dare I say personality, weren't feasible or fast enough on the available local hardware. I'd be interested in other people's experiences here.
So yes, it's early days, and there are potential security issues, but it's an easily accessible glimpse of how valuable AI assistants could be. (PiClawdius was my pun.)