Providers, models & reasoning effort

Choose who does the thinking and how hard: pick Claude Code, Codex, Kimi, or Google Gemini, pick a model, and set the reasoning effort that fits the task.

Switching provider, model, and effort between messages

Where the controls live

The provider, model, and effort controls sit one row below the composer, so you can change them between messages without leaving the conversation. Nothing you pick here is permanent — every control is per session, and switching costs one click.

Choose your provider

The provider picker chooses which agent runs the session: Claude Code (Anthropic), Codex (OpenAI), Kimi, or Google Gemini. They run locally on your machine through the provider CLI you installed and signed in, or via your own API key. Pick the provider that is ready for the task and keep another configured provider as a second opinion. The compact readiness dot always tracks the provider that is currently selected.

Pick a model

The model picker lists the models your provider exposes. Larger models reason more carefully and take longer; smaller ones turn quick edits around faster. Model lineups change with releases, so the current list — with defaults per surface — lives in the Models reference, not here.

Check your credits without leaving the session

When the provider is Axelera AI, a Credits button sits in the same row. It carries no number while an agent works. Click it and the card shows the credits you can spend right now and a gauge beside them: full when your credits are untouched, emptying as you spend, measured against everything your account has had available this month, which is what is left plus what you have already spent, so it only ever falls as you spend and fills again when the month turns or a new grant arrives. Under the gauge the card says how many of the credits this month put within reach you have used and how many are left, and which of them are monthly credits with the date those renew. One-off credits are named apart from that renewal because none of them renew: your welcome credits never expire, and the others last until their own expiry date. The button takes a highlight when what is left will not cover a typical agent task on the selected model, and again when the account is out. Local agent providers spend no credits, so the button does not appear for them.

See requested and running models

The three exact status forms are Requested <exact ID> for your selection, Running <exact ID> only after the active CLI reports its observed identity, and Running model not reported when the CLI is running but supplies no model identity. The display never turns a requested model into an observed claim.

Reasoning effort levels

Reasoning effort is the control most people never touch, and it is worth touching. It does not change which model runs — it changes how long the model thinks before each action.