Browse docs
On this page
Models
Connect a model provider so your agents can reply. Each organization manages its own credentials and billing.
New organizations start with an OpenAI connection that has GPT-6 Luna enabled and assigned to the starter agent, so chat works right away. To use your own OpenAI account, edit that connection and enter your key. To use another provider or a gateway, add a provider with an API key that can access the model you want.
Add a provider
- Open Models and choose Add provider.
- Enter a name, select a provider, and enter its API key. Keep the default API URL unless you use a compatible gateway.
- Enable the models you want. Use the catalog defaults where available.
- Choose Save provider, then select a model under Test model and run the test. Wait for a successful reply. The request may incur provider charges.
- Choose Set up an agent.
Provider formats
Leave the version path off Anthropic and Gemini API URLs. Their clients add it. The provider is fixed after saving. Create another connection to change it.
Keys are encrypted. Leave the key field blank to keep it, or enter a replacement to rotate it. Changing the API URL requires re-entering the key.
Prices and token limits
Catalog prices refresh daily. Use Override catalog defaults for a custom model or different rates. Enter 0 explicitly for a free model. Turning overrides off discards your custom values.
Verify the initial 4,096-token output maximum with your provider. Configured token bounds describe the model, but chat does not currently enforce them as request caps. To control token use or model spend, configure Usage limits. Spend limits depend on having a price for each model.
Assign models to an agent
- Open Agents and choose an existing agent or Add agent.
- Select one or more Models. Each choice includes its provider name, so the same upstream model can appear through multiple connections.
- Choose the Default model, then save the agent. New chat runs use that default through its provider's API URL and key.
Remove models from their agents before disabling them or deleting their provider. See Agents for other settings.
Troubleshooting
- If a model request fails, check the provider's API URL, API format, key access, and exact model ID. Tenant users see only that the assistant is unavailable, so the provider's own error never reaches them.
- If saving refuses the API URL with
Use a public HTTPS API URL. This server does not allow HTTP or private network endpoints, the deployment accepts only public HTTPS endpoints. A self-hosted deployment's operator can allow private endpoints in Configuration. - If the stored key cannot be read, enter it again in the provider editor and save.
- If an agent has no model choices, enable a model on a provider first.