Models and AI keys
Each agent picks its own provider and model on its Editor tab. Who pays for its model calls depends on your organization’s AI key and your plan.
Provider and model
Section titled “Provider and model”An agent runs on one of two providers.
| Provider | Models | Default model |
|---|---|---|
| Vercel AI Gateway | The gateway’s live list of chat models | anthropic/claude-sonnet-5 |
| Workers AI | Cloudflare’s text generation models | @cf/moonshotai/kimi-k2.7-code |
Vercel AI Gateway is the default for a new agent. The Model picker searches by name and groups models by the company that makes them. Its first row, Default, runs the provider’s default model. Switching the provider resets the model to that provider’s default.
The agents list and the chat’s message box both show each agent’s provider and model. A saved change applies to every thread at its next message, and to the next workflow run.
Thinking before each reply
Section titled “Thinking before each reply”Thinking before each reply sets how much the model reasons before it answers: Off, Minimal, Low, Medium, High or Extra high. New agents start at Medium.
The setting applies to reasoning models from Anthropic and OpenAI on Vercel AI Gateway. Other models ignore it. In chat, the model’s reasoning shows folded under Thinking above its reply.
Long conversations
Section titled “Long conversations”When a thread grows close to the model’s context window, the agent condenses its older turns into a summary. This happens only for models whose context window the catalog states.
Who pays for a turn
Section titled “Who pays for a turn”Every message in a chat and every Agent step in a run is a turn of one or more model calls.
| Situation | What the turn runs on |
|---|---|
| Vercel AI Gateway, your own key set | Your key. Vercel bills you, and nothing counts here |
| Vercel AI Gateway, no own key, Pro plan | Our key, counted against your monthly AI allowance |
| Vercel AI Gateway, no own key, Free plan | Nothing. The turn is refused |
| Workers AI, any key | Our Cloudflare account, only on a plan that includes AI |
Pro includes $5 of AI a month on our key. Free includes none, so agents on the Free plan run only on your own Vercel AI Gateway key. Your own key never applies to Workers AI. See Plans.
Add your own key
Section titled “Add your own key”Only owners and admins see the AI key card.
- Create a key in your Vercel AI Gateway account.
- Open Settings.
- In the AI key card, paste it into Vercel AI Gateway key.
- Press Save.
The card then says your agents use your key and shows its last four characters. The key is stored encrypted and never shown again in full. To go back, press Use the platform’s key.
The key is read at the start of every turn. A new key, or a removed one, applies to the next message in every thread and the next Agent step in every run.
When the AI allowance runs out
Section titled “When the AI allowance runs out”AI spend is the estimated dollar cost of each model call on our key, as the gateway reports it. It counts per calendar month in UTC. Agent chats, Agent steps and the assistant all count. If an operator granted your organization AI credit, spend above the allowance draws on it.
Once the month’s allowance and any credit are spent, turns on our key and on Workers AI are refused with a message that says so. A chat shows it under the conversation. A workflow’s Agent step fails with it. Add your own key in Settings to keep going, or wait for the next month.
The Billing page and the dashboard show what is left this month. See Allowances.
The assistant and Improve with AI
Section titled “The assistant and Improve with AI”The assistant, including Improve with AI in the agent editor, is on the Pro plan and always runs on our key. Its turns count against the AI allowance whether or not you have set your own key.