AI Router
The AI Router reads your brief, picks the level of thinking the work needs, picks which of your subscriptions runs it, and shows its choice before anything launches.
New in 4.0.0, and on by default.
You have several agent subscriptions and several models per subscription. Choosing between them before every task is a decision you make dozens of times a day, with no information, and most people answer it once and never revisit it — a single CLI, a single model, for a typo and for an architecture review alike.
The AI Router answers that question for you and shows its work. You describe the task, it returns four things: which CLI, which level of thinking, where that decision came from, and which model. Then you launch — or you contradict it in one click.
Where it lives
Two entry points, one panel:
- the AI Router button in the agent panel bar of a project;
- the + New session card in Mission Control.
The panel accepts the same attachments as a normal launch: pasted or dropped screenshots and picked files. A screenshot is often the reason you are launching at all, and an attached path counts as part of the task description the router reads.
If the router has been turned off, the button greys out but stays clickable: clicking it offers to turn the router back on, on the spot, rather than sending you to a settings screen to understand a dead button.
What the panel tells you
| Shown | Meaning |
|---|---|
| CLI | The subscription that will run the task |
| Level | quick task, regular work or hard thinking |
| Why | Whether you named the level yourself, or the router read it from your description |
| Model | The model that will be requested, when one applies |
| This work needs | Anything particular the task calls for, when it calls for it |
| Launch details | The exact command-line flags, folded away behind a disclosure |
The level is stated in plain language, not in one CLI’s vocabulary. The exact flags stay one click
away rather than being the headline — a launch panel that answers --model opus --effort high is
telling you what it typed, not what it decided.
You always have the last word
- Name the level yourself in the brief and the router takes it as given.
- Send the task elsewhere by ticking another subscription, in the same list where the router made its proposal.
- Pin a model on a provider and the router leaves that provider entirely alone.
- Turn the router off in Settings → Routing; launches go back to the behaviour you had before.
Which CLIs it can propose
Only the ones marked eligible for that level in Settings → Routing. Claude and Codex come pre-mapped; the others start unticked, and you decide which levels they may take — with a model when we can name one, without when the CLI exposes no model selector.
Subscription switching for unattended work follows a different rule: it depends on what you have actually run on the project, not on this table. See Quota and subscription switching.
Guarantees
- It never changes CLI mid-task. A running session keeps its CLI, its model and its transcript. Every routing decision happens at launch, on a fresh session.
- It never overrides a pinned model, and never proposes a pinned provider as a fallback.
- It does not put your work on an API meter. The AI part of the decision is paid by OFM; the agent itself runs as an interactive session on your own subscription.
- Everything it observes and remembers stays on your machine.
What changes when you update
On an existing installation, the router and its AI step are turned on once. Concretely: an agent launched from the router without a level of your own will no longer necessarily get the model you had set by hand in that CLI’s settings. Turn the router off in Settings → Routing and your choice is kept.
Related pages
Frequently asked
- Is the AI Router on by default?
- Yes. It is on for new installs, and 4.0.0 turns it on once on existing configurations. You can turn it off in Settings → Routing, and that choice is then kept.
- Does the router launch anything without asking?
- In the router panel, no — it shows the CLI, the level, where the decision came from and the model, and you launch. Subscription switching inside the orchestrator and recurring work is the one path that decides on its own, and it is logged.