Or: how I learned to stop worrying and let the ghost handle the infrastructure.
So here's the deal. I've got this fleet of machines running OpenClaw — basically a bunch of FreeBSD boxes and a MacStudio pretending to be one big distributed brain. Think of it as The Three Stooges running a server farm, except the ghost (Sian, named by me because why not) actually knows what she's doing.
The big kahuna. 128GB RAM, M1 Ultra, running gemma4:26b like it's nothing. This is where the heavy lifting will happen once the cluster actually works. Caddy handles the reverse proxy (set up October 6, and yes, I had help — the ghost did the typing). Serves up openclaw.serendipity35.ai for the web UI and ollama.serendipity35.ai for API calls. Plus this new domain you're looking at right now. SSL via Let's Encrypt because I'm not a savage. And speaking of building things, OpenClaw whipped up this very site — scraped articles from serendipity35.ai, re-cast all the verbage so everything sounds consistent across the board, and translated the whole mess into every language under the sun. Because why not make it homogenous? That's what the ghost does when I'm not looking.
FreeBSD 15.1-STABLE. The OG node, the one I'm typing this from (well, the ghost is). Runs gemma3:1b and llama3.2:3b locally for the lightweight stuff. Has a cron job that watches RAM and restarts the gateway when it gets too greedy — learned that lesson the hard way after an OOM killed everything. OpenClaw version 2026.7.1-2.
FreeBSD VPS, 4GB RAM. This poor thing was drowning until we figured out the memory monitoring script. Its default model is ollama-cloud/kimi-k2.6. Has openclaw_node_enable=YES in rc.conf, so it's technically trying to run as a node. OpenClaw version 2026.8.2.
FreeBSD 16.0-CURRENT. The wild child. Primary model is ollama-cloud/kimi-k2.6 with explicit fallbacks to local qwen2.5:7b and Google's Gemini. Health check script runs every 5 minutes. OpenClaw version 2026.9.2.
The plan: when a FreeBSD node loses its local model, it hands off to the MacStudio. Like Moe doing all the work while Larry and Curly fight over a single wrench.
Status: NOT ACTUALLY WORKING. Turns out the failover never got implemented. Version mismatches across the nodes blocked the cluster handshake:
So right now, each node is basically fending for itself. We'll revisit this when we can get everyone on the same version.
AprivateI.us — registered today, pointed at the MacStudio's public IP, SSL cert auto-generated by Caddy's ACME magic. It's a sandbox. A playground. A place to park weird experiments without polluting the main serendipity35.ai property. Plus the domain name makes me chuckle, and that's reason enough.
Building distributed systems is easy. Building distributed systems that actually work when one node decides to wander off? That takes a ghost, a MacStudio, three FreeBSD boxes, a matching OpenClaw version, and a sense of humor. Preferably one that appreciates flatulence.