One client, your whole lab: the tour

Level ReadingEstimated time 10 minA word you don't know? Glossary

Picture this. You are on the sofa. On your phone you open one app. You ask a coding agent to fix a test on the server, you watch it work line by line, you approve its command with a tap. A notification tells you the image render you ordered an hour ago is ready. Another says the guard machine banned an address scanning your proxy. All of it on your hardware, with no monthly bill per token.

Two screens, one back end. The phone app and the desktop client never talk to an agent directly: they go through a small API that reads the fleet, and a rooms controller that keeps every conversation. Reads run alone; any action waits for your tap on a confirmation card.

Proven here, every day. Not a mock-up: this is the setup this site was built with. Each piece below has its own step-by-step tutorial, with the commands, the numbers we measured and the traps that cost us an evening.

What one client lets you do

Code on every machine. Open a session with OpenFox on any of your machines, follow it live, step in, approve. Long tasks go to Kimi Code, delicate ones to Claude Code, everyday ones stay local and free.
Keep the thread. Rooms survive a closed app, a lost signal or a change of screen: start on the laptop, finish on the phone.
Create. Images, video shots, 3D objects with ComfyUI on your own cards, ordered by an agent that fills in a tested workflow.
Watch over the lab. The guard machine (proxy, fail2ban), errors, backups, which model sits on which card: all in one summary, alerts pushed by ntfy.
Decide, in one tap. The local orchestrator reads everything by itself and proposes; nothing acts without your tap on a card (see Inari).

Why it stays simple

The concept is ambitious; the building blocks are not. Three rules keep it that way:

Small pieces, one job each. An API that reads, a controller that keeps conversations, agents that code. Each fits in your head, each can be replaced.
The model proposes, code decides. Verdicts, merges and actions are plain programs or your tap, never a model's mood (see BMAD with small models).
One gateway to every model. Change the model, the card or the vendor; nothing else moves (see the model gateway).

Where to start

You do not build it in one go. Start with the gateway and OpenFox: one evening, and you already have a local coding agent. Every next tutorial adds one brick that works on its own.

Want it set up for you, or adapted to your own lab? Get in touch.

Found this useful? Tell us what to measure next.