Like a lot of other people, I woke up Saturday deciding that I had to run OpenClaw.
I’ve spent big swaths of my career either being the least technical technical person or the most technical business person in the rooms I enter. I like it. The emergence of generative AI has been a boon for people like me— I am used to working on the edge of code blocks, arguing with computers, and cajoling page loads.
Now it was time for me and my terminal to create my own agent.
OpenClaw has some well-documented security concerns (here’s a funny / serious Cisco critique), so I was scared/ terrified/ unqualified to run it from my actual computer.
I thought about buying a dedicated box and hosting it at home (there’s been an OpenClaw-centric run on Mac minis), but that introduces a whole set of operational tasks I don’t want or need: power, ISP hiccups, router config, and the security overhead of exposing something on my own network.
A SaaS tool like MindStudio is 100% my speed, and I am going to develop more on that soon, but I wanted to experience the control of installing and configuring openclaw myself. Time for me to run shell commands like I’m Root in Person of Interest.
So I detoured into the safer, controlled option: a hardened DigitalOcean droplet. It gives me a clean, isolated environment, decent defaults on security, and a clear boundary between “my normal computer” and “the machine that can actually do things on my behalf.”
Here’s the processes, tooling, and results over ~10 hours spread over two days to create PrimO, our household’s first agent:
Core AI + model layer
- OpenClaw – the self‑hosted personal AI agent
- Anthropic Claude models – Haiku, Sonnet, Opus variants used as primary
- Claude API key – Anthropic credentials the agent uses to talk to Claude
- Google Gemini models – Gemini 3 Pro / Flash used as fallback models when Claude complains that I don’t give it enough money
- Gemini API key – Google AI Studio / Gemini credentials
- OpenRouter – model router/provider wired in as a unified backend
Infrastructure + OS layer
- DigitalOcean droplet – hardened VPS that isolates the agent from my laptop
- Terminal / shell – my glorious command‑line front end
- systemd – supervising the gateway (systemctl restart openclaw-gateway.service).
- lsof / kill – used to find and terminate the old clawdbot-gateway
OpenClaw runtime + files
- OpenClaw CLI – all the openclaw … commands, including:
- openclaw status, openclaw status –deep
- openclaw doctor, openclaw doctor –fix
- openclaw onboard, openclaw configure
- openclaw gateway stop/start/health
- openclaw config get/set/unset
- openclaw models list/status/auth login
- openclaw channels status/login
- OpenClaw gateway – the WebSocket gateway that actually runs PrimO.
- OpenClaw workspace – /root/.openclaw/workspace/, where I put big files for the agent to review
- Nano – in‑server text editor for openclaw.json and other configurations
- scp – used to upload files from my laptop into the workspace
- 1Password as a manager of API keys etc
Messagiung & interaction layer
- WhatsApp – main way we talk to the agent day‑to‑day
- Dedicated phone for the WhatsApp bot – the line OpenClaw is logged into— a shared phone we already had for other purposes
- Personal WhatsApps accounts – only SL and I are allow‑listed / paired to PrimO
What’s next? I realized through all of this that the command line is not where I live. Time to go back to connecting LLMs via zapier, make, and monday— I orchestrate.