Routing modes
Each AI has its own mode, set in its editor. The mode is the consent: it decides, up front, what routing may do.
| Mode | What it may do |
|---|---|
| Auto - Offline Only | Never touches the internet. Picks among your installed models by how each runs on this machine - full speed, GPU + RAM, or too large - and among the models on a server you run, when you have one. The mode name is the promise. |
| Auto - Online and Offline | Answers with a frontier online model by default and keeps a question on your device, or your own server, when a model there is better for it, by the rules. Health questions never go online; attachments stay home unless you say otherwise. |
| A specific model | No routing at all - that model answers everything. A pinned offline model stays offline, full stop. A model on your own server can be pinned like any other. |
What "offline only" means in practice
An offline-only AI makes no network requests for a reply, ever, beyond your own machines. In project sessions it stays off the web too - the agent's web search is refused, and the rail says so, rather than quietly routing around the promise. The app's own routing check holds it to that: an offline-only AI sends nothing online on any dial.
What "online and offline" means in practice
Online is the default for an ordinary question, not the exception. The rules decide which online model takes it - everyday, hard, or the live web - and the one dial decides how much stays home: Frontier-first, Balanced, or Local-first, where online is for the live web alone. The mode is the consent; the dial is the leaning. The reply always names where it ran.
Your own server is your own hardware
A server you run yourself - another computer with llama.cpp, a vLLM box, a Mac in the next room - is part of your offline models once it is connected in Settings › Engines. There is no separate mode for it. In either Auto mode the server takes a question when it holds the clearly stronger model for it, and stays out of it otherwise; the reply says "your server" when it answered. A server that stops answering is set aside for a minute and work carries on here.
Earlier versions had a third choice, Auto - My Hardware. It is gone: an AI still set to it behaves as Offline Only, which is what it promised.