I wanted to find out whether I could fit an AI agent into my maker workflow without needing expensive hardware. So I took my oldest Raspberry Pi, a Raspberry Pi 3 Model B Plus with just 1 GB of RAM, and tried to run Hermes on it.
I did not expect this to be easy. The board is small, low power, and very limited, which made it a good test case. If it worked here, it would show that I do not need cutting edge hardware just to begin using AI agents in practical projects.
How I set it up
I started with Raspberry Pi Imager and flashed Raspberry Pi OS Lite 64-bit to a USB drive instead of an SD card. The Pi had already been configured to boot from USB, so I could keep the setup fully headless and join it to the network with Wi-Fi and mDNS.
Once it was online, I connected over SSH and ran the Hermes install script from the Hermes Agent site. The first setup pass did not go perfectly because I hit Ctrl+C and interrupted it, so I had to restart the configuration with Hermes setup.
Choosing the model and chat app
For the language model, I paired Hermes with OpenAI Codex. The setup also supports other providers, including local models, but I chose a cloud model for this test so the Pi did not have to do the heavy compute work itself.
I then set up Telegram as the chat interface. I used manual setup, created a bot, and locked it down so it would only talk to me. That gave me an easy way to send messages to the agent from my phone.
Fixing the identity problem
I also gave the agent a name: Gort. I set an avatar for it and tried to configure its SOUL.md so it would remember its own identity and purpose.
That part was messy. I found that I was updating a SOUL.md file in the wrong place, so the agent kept forgetting who it was. It only started behaving properly when I told it to update the system folder version instead.
Putting it to work
After the setup was stable, I gave Gort the Universal Embedded Workbench repository and asked it to configure a skill for that tool. I wanted to see whether it could learn its way into a real maker task instead of just chatting.
My test job was simple but useful: write an ESP32 sketch that flashed an LED in a countdown pattern from five down to one, pause for five seconds, and then repeat. I also wanted it to fit my PlatformIO project layout so I could open it properly in Visual Studio Code later.
Gort wrote the code and tried to upload it to the workbench. The first attempts failed, but it noticed the error, kept track of what went wrong, and tried again without me telling it to do that. After another pass, the upload worked.
What surprised me
The most surprising thing was that this all worked on a Raspberry Pi 3 with only 1 GB of RAM. Even Telegram voice dictation worked, although it was slow on this hardware.
That makes me think the real limit is not whether AI agents can run at all, but how carefully I set them up and what I ask them to do. A small, low power machine is enough for me to start experimenting.
My next step is to write proper specification documents and use more frontier models to plan and execute larger programming jobs in a token efficient, inexpensive way.