OpenClaw on a Raspberry Pi
The pitch is hard to resist. Take a Raspberry Pi you probably already own, install OpenClaw on it, and you get an always-on personal agent for the one-time cost of a takeout dinner. No monthly hosting fee, no laptop that has to stay awake. A little box on a shelf that reads your messages, watches your inbox, and does things for you while you sleep. I was in immediately.
The setup was honestly the easy part
I expected the install to be the painful bit. It wasn't. The official docs have a proper Raspberry Pi path, and the whole thing is maybe thirty minutes if your SD card cooperates. The Pi is only the gateway; the models run in the cloud via API, so even a modest Pi 4 handles it. The short version:
# flash Raspberry Pi OS Lite (64-bit), enable SSH, then:
sudo apt update && sudo apt upgrade -y
sudo apt install -y curl git build-essential
# Node 24 via NodeSource (the apt version is ancient)
curl -fsSL https://deb.nodesource.com/setup_24.x | sudo -E bash -
sudo apt install -y nodejs
# OpenClaw itself, plus the onboarding wizard
curl -fsSL https://openclaw.ai/install.sh | bash
openclaw onboard --install-daemon
# keep the gateway alive when you log out
sudo loginctl enable-linger "$(whoami)"A few things the guides say that turned out to matter: use the 64-bit OS (not optional), add a swap file if you're on a 2 GB Pi, and prefer API keys over OAuth for a headless box. That last one I'll come back to, because it turned out to be foreshadowing.
Texting your own agent never stops being fun
The genuinely clever part of OpenClaw is the channels. You don't open a web app to talk to your agent; you message it on Telegram or Discord like it's a person. Sitting on the couch, firing off "summarize whatever came into my inbox today" to a Telegram contact, and watching a computer on my shelf actually do it: that felt like the future. When it worked, it really worked, and I would grin at my phone like an idiot.
Then the bills and the lag showed up
Two problems arrived together. First, the Pi experience was slow and finicky at times. Not broken, just laggy enough that you feel the difference between "instant assistant" and "little computer doing its best." Second, and worse: an always-on agent that thinks with Claude or ChatGPT class APIs eats tokens at a frightening rate. Every message, every email summary, every background check-in is a metered API call, and an agent that never sleeps never stops spending. My experiment was suddenly a subscription with extra steps.
So I did what everyone does at that point and switched to open-source models. This was four or five months back, and I'll be honest: at the time, they weren't doing a whole lot for this kind of agentic work. Tool calls came back malformed, multi-step tasks fell apart in the middle, and the agent got noticeably dumber the moment I stopped paying per token. The field moves fast and I'm sure this has already aged badly, but back then the choice was pay a lot for good, or pay nothing for frustrating.
The Splitwise agent, and where it all unraveled
The setup that made it all feel worth it, for a while, was house-keeping automation. I had the agent going through my emails and summarizing them, and I hand-wrote a small skill for Splitwise so I could forward a bill and have it split with the right people. Then I got greedy and wired in the follow-on steps: after adding an expense to Splitwise, log it in Google Sheets, then deal with the related email in Gmail.
That's where things got weird. The agent needed frequent re-authentication with Google, and every time a token expired the whole chain quietly fell over. Splitwise would get the expense, then the Sheets step would fail, or the Gmail step would stall, and the agent would start acting strangely instead of failing cleanly. I re-authenticated a couple of times, each time by juggling OAuth flows onto a headless machine, and then one evening I realized I was maintaining the assistant more than it was assisting me. I gave up.
In hindsight, a Mac Mini might have made this easier. A real desktop session to complete OAuth flows in, more headroom, fewer of the small ARM-flavored papercuts. But Mac Minis were getting expensive then and still are, and "buy a nicer computer" was not the experiment I signed up for. The Pi was the point.
Would I do it again? Absolutely
Here's the thing: none of this reads like a complaint to me. The setup was smooth, the idea of meeting your agent inside the apps you already use is genuinely great, and the direction is obviously right. I hit the rough edges of a very young project at a very early moment, running it on the cheapest possible hardware, with models that weren't ready to be frugal. Every one of those constraints has been improving monthly since.
I'm also looking forward to trying Hermes, the open-source agent from Nous Research that everyone's talking about; the premise of an agent that writes its own skills after each task is the kind of thing my Splitwise experiment could have used. The little box on the shelf is patient. And despite the token bills, the re-auth loops, and the evening I rage-quit, the whole experience was, in itself, an enjoyable one. Some projects pay you in output. This one paid in fun.