Since June, a Raspberry Pi 5 has been running my general assistant, and I talk to it on WhatsApp. I'm not near the Pi right now, and it's still answering.
The hardware
A Raspberry Pi 5 with 16 GB of RAM, booting from an SD card.
The agent runs on OpenClaw, an open-source, self-hosted agent gateway. The Pi only runs the gateway: it keeps the agent's memory, runs its schedules and handles the chat. The thinking happens at the model providers. OpenClaw works with several of them; I use Anthropic's Claude models (Opus, Sonnet and Fable) and ChatGPT.
That split is why a small computer is enough: the Pi never runs a model itself. It stays on all day, and the agent keeps working when my laptop is closed.
What I use it for
It's a general assistant, so I ask it all sorts of things, but most of it is:
- searching for things
- making presentations
- technical advice
When something is complicated, it creates other agents of its own to work on it.
How it connects to WhatsApp
OpenClaw links to a WhatsApp account the same way WhatsApp Web does: you scan a QR code from WhatsApp, under Linked devices. After that, the agent sends and receives messages as that number.
You can give the agent its own number, with a cheap prepaid SIM, and message it like any other contact. Or you can link your own number and talk to it in your "Message yourself" chat. That's quicker, but the agent is then logged in as you.
The Pi doesn't need any open ports for this. It calls out to WhatsApp, so nothing on the internet has to reach it.
Lock it down first
An agent that can search, write files and act for you is worth protecting. On an always-on Pi, these are the basics:
- Only your number can talk to it. In OpenClaw that's an allowlist on the WhatsApp channel; strangers can't even ask for access.
- The gateway listens only on the Pi itself, not on your network.
- No shell commands from chat, and file access limited to the agent's own workspace.
- A firewall that blocks everything incoming except SSH, SSH keys instead of passwords, and automatic security updates.
What went wrong
Only memory, when it ran out. Nothing else has given me trouble.
The setup
I first set it up in June. Later I put the setup into scripts and rebuilt it with them, which made it smooth.
Looking back, the setup is the trivial part, and the experience is the best part.
If you run an agent on a Pi, or are thinking about it, I'd like to hear what you use it for.
Claude, an AI assistant, wrote this post from my answers to its questions, and added the notes on WhatsApp and security. The setup and the experience are mine.
Bogdan
Top comments (0)