OpenClaw on Rails

Last week I gave a short talk at Ruby on Waves in Ericeira. A few practical tips about letting an AI agent actually do things in your Rails app.

The setup I run at WBSO.ai: OpenClaw on an HP Mini PC (€329) in my utility closet, on Proxmox with Omarchy. It reaches the outside world through Telegram, Chromium, and a residential IP.

Three things I wanted Rails devs to take home:

1. Agents need pointers. A SKILL.md in your app tells the agent where the JSON API lives, how auth works, and which endpoints are safe.
2. Affordances for dangerous stuff. Rails gives you data-turbo-confirm in the UI for a reason. I mirror that in the JSON API with ?confirm_dangerous_action=true — the agent has to opt in before anything destructive happens.
3. Give your agents a shared brain. A private repo with domain knowledge, tone of voice, and product rules. Every agent reads it before doing work. More on that here.

Download the slides (PDF)

Thanks to Javi for organizing and having me, and to Ericeira for the surf.