One AI Server on a single Windows, macOS, or Linux machine, serving only that machine, is the configuration every AI Suite desktop app creates for itself automatically <a href="#ref-1">[1]</a>. Running it deliberately shows what that default actually is. The daemon binds <code>127.0.0.1</code> only, so nothing else on the network can reach it; it needs no keys and no licence, because loopback serving is free; and every app on the machine shares the one daemon, which means one model in memory rather than a copy per app <a href="#ref-1">[1]</a>.
Nothing on the network can reach it
Loopback is not a firewall rule you might forget to add. The socket is bound to the local interface, so the endpoint does not exist to any other machine <a href="#ref-2">[2]</a>. That is the default posture worth understanding before anything is deliberately networked: know what runs locally, then decide what to expose.
One daemon, one model in memory
Because apps on the same machine discover the daemon automatically through its <code>server.lock</code> file, opening a second app does not load a second copy of the model <a href="#ref-1">[1]</a>. The chat app, the notes app, and the PDF reader all attach to the same running server and the same warm model.
Stub for CI, real for work
The <code>stub</code> engine starts instantly with no downloads, which suits CI agents that need a local AI endpoint for tests; the <code>real</code> engine downloads a model on first use for actual inference <a href="#ref-2">[2]</a>. A health check and an OpenAI-shaped chat completion confirm it is up. The built-in dashboard shows the mode badge and the request you just made, no key required on the same machine.
Nothing to configure
There is nothing else to set: no keys, no licence, no network policy for the single-machine case <a href="#ref-1">[1]</a>. It runs on the laptop you already own, fully offline. When you are ready to serve more than one machine, the same daemon binds the LAN behind a key.