Home & lab · Feb–Mar 2026 · with Delivered B

A Raspberry Pi and an agent experiment

A Raspberry Pi in a drawer, a lightweight agent framework and a cheap model API. What I set up, what broke, and why I retired it.

The idea

I had a Raspberry Pi 3 Model B in a drawer and wanted to see how far a small always-on box could go as a home for an AI agent: something that could be messaged, run tasks and sit quietly on the network without taking over a computer I use. PicoClaw is a lightweight agent framework aimed at exactly that kind of small hardware, so it was the thing to try.

Step 1: get into the Pi without passwords

Everything else is easier once login is key-based. Make a key on the Mac, copy it across once, then give the Pi a short alias:

# on the Mac
ssh-keygen -t ed25519 -C "mac-to-pi"
ssh-copy-id pi@<pi-address>          # one-time, uses the password

# ~/.ssh/config
Host pi
    HostName <pi-address>
    User pi
    IdentityFile ~/.ssh/id_ed25519
    IdentitiesOnly yes

Along the way I caught a one-word typo in the SSH config. When login misbehaves, ssh -v pi shows which config lines and which key SSH actually used, which is the quickest way to find that kind of mistake.

Once key login works, turn passwords off so the Pi can't be guessed into:

# on the Pi, after key login works
sudo nano /etc/ssh/sshd_config
#   PasswordAuthentication no
sudo systemctl restart ssh

Test a second login from a new terminal before closing the first one. If the key setup has a mistake, you still have a way back in.

Step 2: choose a model provider

The Pi only runs the agent. The thinking happens at an API provider, so the comparison was about price and flexibility, not hardware. I went through the direct options and the aggregator OpenRouter, which puts many models behind one key and one bill. For an experiment, the winning combination was OpenRouter with DeepSeek V3: capable, and cheap enough that a runaway loop costs cents rather than dollars.

Step 3: the two settings that bit me

• An empty model setting. The agent wouldn't work until the model setting, which had been left empty, was filled in. The fix was one line. The lesson is that “it starts” and “it is configured” are different states, so send one test message before changing anything else.
• A token limit that was too low. A token limit in the config was too low for real replies, and raising it fixed the problem.

Hardware, and where it ended

A Pi 3 is enough to run the agent, which is the point of the framework. I also asked about mini PCs for the larger OpenClaw agent, which wants more memory, and looked at a Beelink S13 Pro. I never bought one.

The working outcome was a configured agent on the Pi and key-based login from the Mac. In October 2026 I retired the project.

What I'd tell someone starting

• Keep API keys out of notes and chats. Put them in the tool's config or an environment variable, and rotate any key that has ever been pasted somewhere it shouldn't be.
• Set a spending limit on the provider account before connecting an agent to it.
• Do the SSH key step first. It makes every later step quicker.

This is entry 03 of AI Lab Work.