John Mullins
Welcome to my blog where I share thoughts on code, hacking, and life.
Welcome to my blog where I share thoughts on code, hacking, and life.
Back in January, two friends and I started experimenting with OpenClaw. We each took a different path. One ran it on a rack-mounted Mac Mini. Another set it up on Proxmox. I went with a Digital Ocean VPS.
That first week taught me a lot about the ecosystem. I built custom tools for notifications and email. The real win was something I called NightShift, a program that would pick up projects defined in Linear and push pull requests to GitHub overnight.
Then I ran the OpenClaw update command and everything crashed.
I spent the next week rebuilding. Docker Compose, a local LLM on Ollama, and several forks of OpenClaw later, I had a clear goal: bring that $130 weekly bill down.
I landed on devstral-small-2:24b with Picoclaw. Running locally on my M1 MacBook meant roughly one minute per request. Fast enough to be useful, slow enough to notice.
Eventually I settled on a shared CPU Vultr instance running Arch Linux, Picoclaw, and OpenRouter models. A small Go binary I felt comfortable modifying myself. The whole setup runs on hardware you wouldn't blink at.
That's the real lesson here. You don't need a powerful machine to run an AI agent. You just need reliable access to an LLM. OpenRouter's dynamic routing handles model selection automatically, and my weekly costs sit around $5.
A few months ago I was skeptical. Now I run my own stack, spend almost nothing, and ship code while I sleep.