Two weeks ago I launched ClawHosters, a managed hosting platform for OpenClaw (open-source AI agent framework). 50 paying customers so far. Here's the technical post-mortem.
The stack: Hetzner VPS provisioning from snapshots, 5-layer Traefik routing with Redis provider, Nginx reverse proxy that kept silently breaking SSE streams, Docker container isolation with a writable-layer commit strategy, and a config migration registry to handle crash loops at 2 AM.
Some highlights from the article:
- Why `proxy_buffering off` alone doesn't fix SSE through Nginx (you need a re-framing buffer)
- Docker's `cp` command silently refuses symlinks — cost me a day
- ZeroTier + `nsenter` for injecting routes into running containers without restart
- Snapshot-based provisioning: ~3 min cold, ~30s with pre-warmed pool
- Per-request LLM billing with included token allowances per tier
Built as a solo developer alongside a full-time job. The whole thing runs in a single Rails monolith.
The stack: Hetzner VPS provisioning from snapshots, 5-layer Traefik routing with Redis provider, Nginx reverse proxy that kept silently breaking SSE streams, Docker container isolation with a writable-layer commit strategy, and a config migration registry to handle crash loops at 2 AM.
Some highlights from the article:
- Why `proxy_buffering off` alone doesn't fix SSE through Nginx (you need a re-framing buffer) - Docker's `cp` command silently refuses symlinks — cost me a day - ZeroTier + `nsenter` for injecting routes into running containers without restart - Snapshot-based provisioning: ~3 min cold, ~30s with pre-warmed pool - Per-request LLM billing with included token allowances per tier
Built as a solo developer alongside a full-time job. The whole thing runs in a single Rails monolith.