Models
A model is the one capability every agent needs and the one nobody thinks of as a capability. Naming a provider is the whole of configuring it.
A list, and mostly just a provider
models:
- id: deepseek-v4-flash # the id doubles as the model when it already is one
provider: deepseek
- id: sonnet
provider: anthropic
model: claude-sonnet-4-6
defaults:
model: deepseek-v4-flash # where every agent starts
It used to be four coupled things — the provider, the model, a placeholder key in the sandbox's environment and a grant naming the provider's host — and any one of them wrong is not a startup error but a turn that dies at the proxy, complaining about something the operator never typed. Where a provider lives, what its key is called and whether the key goes in a bearer header or one of its own are facts about the provider rather than decisions anybody gets to make, so they are not written out.
A provider nothing here knows still works by saying the two things the table would have said: a host and a keyEnv. The key itself is never in this file and never in the agent — keyEnv names a variable of the control plane's own environment, and the grant each model produces is what writes it onto the request on the way out.
/model is a choice among what exists
▸ /model flash deepseek/deepseek-v4-flash (this one) /model sonnet anthropic/claude-sonnet-4-6 /model gpt-5 openai/gpt-5 (no OPENAI_API_KEY) ╭──────────────────────────────────────────────────────────────────────╮ │ > /model │ ╰──────────────────────────────────────────────────────────────────────╯ ↑↓ model ⏎ choose ^C quit
Each row says the two facts about a model that are not its name — whose it is and what they call it — plus the one that decides whether the next turn answers at all. A model this plane holds no key for is offered and marked, not hidden, because it is configured and the missing half is a key you can paste in two panes over. Typing narrows the list against the id, the provider and the provider's own name together, so /model anthropic finds the one called sonnet.
Every model on that list is already reachable by every agent — configuring one is what granted it — so moving between them changes what a turn costs and how good it is, and changes nothing at all about what the agent can get to. That is what makes it a command rather than an edit and a restart, and it is why an agent is allowed to ask for it.
Nothing is recreated to do it. The container was started holding a placeholder for every provider this knows, and the runner asks what to think with at the start of every turn — so a switch lands on the next turn, and a turn already running finishes on the model it was handed when it started. That last part is said out loud, because the change looks instant and is not.
Both halves are a list on the config screen
│ config │ │ │ │ holds from the next turn — nothing restarts. │ │ │ │ providers │ │ ● deepseek DEEPSEEK_API_KEY flash │ │ ○ anthropic ANTHROPIC_API_KEY sonnet │ │ ○ openai OPENAI_API_KEY gpt-5 │ │ ○ groq GROQ_API_KEY no models │ │ │ │ models │ │ ● flash deepseek from the file │ │ ○ sonnet anthropic from the file │ │ ○ gpt-5 openai added here │ │ + a model │ │ ╭────────────────────────────────────────────────────────────╮ │ │ │ ANTHROPIC_API_KEY no key, refused at the proxy │ │ │ ╰────────────────────────────────────────────────────────────╯ │
A model is three lines of configuration and one exported variable, and the variable is the half that is not in the file — so it is the half that gets forgotten. The failure that produces is a plane that is running and configured and refused at the proxy, with turns dying over a host nobody typed. ● is something this plane can use right now and ○ one it cannot, the same mark the agents column uses.
⏎ on a key takes it, masked as it is typed and never shown again. It goes to the plane over the same socket a shell does, and for the same reason: holding that socket is what makes somebody the operator, and a key arriving by webhook would be a stranger paying with your account. A key pasted here holds on the next turn, with nothing restarted and nothing redeployed; an empty line takes it back, and the question goes to the plane's own environment again.
+ a model asks the providers rather than asking you
│ 3 on offer │ │ › gpt-5-mini openai │ │ gpt-4o-mini openai │ │ o4-mini openai │ │ ╭────────────────────────────────────────────────────────────╮ │ │ │ model openai mini │ │ │ ╰────────────────────────────────────────────────────────────╯ │
Being handed a key and then asked for a model name is being asked for the one fact the key just made the plane able to look up. So every provider it holds a key for is asked what it answers to, all at once, and what comes back is a list to arrow through. Typing narrows it against the provider and the id together and in any order, so openai mini gets there without remembering which of gpt-5-mini and gpt-5-nano was the one.
A provider that would not answer is named under the list instead of being counted as having nothing, since an empty list is the shape both a wrong key and an empty catalog arrive in. Writing one out by hand still works and has to — three words, a name, the provider it thinks on and the provider's own name for it, so sonnet anthropic claude-sonnet-4-6 is taken as typed.
What the file declared is there to be read and not to be changed. from the file is a row this screen will not shadow and will not drop. Everything given here lives in a store beside config.yaml and never in it, so a redeploy brings back what was written there and what was typed here survives the redeploy on its own. This is the one screen where the keyboard grants rather than pays, and that is deliberate: reaching it means holding the plane's control socket, which is the whole of being the operator here.