The order
- Observability — once you have services worth watching. Metrics, logs, and traces so you’re never guessing what your own box is doing. Shipped here as a clearly labeled optional module:
optional/observability/. Read its README before you turn it on; the point is that you don’t start with a monitoring stack. - More tools — when you hit a wall a new MCP server removes. Add the one that unblocks you, not a catalog.
- A second agent / model — when one model’s weakness starts costing you real time. The gateway already makes this a config change.
- Databases — when something needs to persist structured state (a real app, a knowledge base with vectors). Run them in Docker, bound to the tailnet.
- A deploy path — when you’re shipping public sites. Serve them through an outbound tunnel (e.g. Cloudflare Tunnel /
cloudflared) so you keep zero ingress: the box dials out, nothing dials in. - GPU, rented by the job — the first time a task genuinely needs a card. Burst to a serverless GPU backend for the length of the job; don’t buy one to idle. See
WHEN-NOT-TO.mdfor the exception (24/7 GPU-bound work).
Things people build on this
- A coding environment any model can drive.
- A private second brain — a knowledge base your agents read and write, on hardware you control.
- A prototype factory with real CI and a real deploy path, no per-experiment cloud bill.
- A research rig that borrows GPU muscle by the minute.
- A self-hosted automation hub — watch a source, summarize, act, notify.
- A place to actually learn infrastructure by running it.