Skip to content

Models

A model connection is a provider plus the credentials to use it. Agents pick their models from these connections, so you can switch models without touching the agent’s prompt or plugins.

  • Cloud: OpenAI, Google Gemini, Mistral AI, Groq, OpenRouter, DeepSeek, xAI, Cerebras and Fireworks AI.
  • Local: Ollama and LM Studio. Point the connection at the server’s URL, like http://localhost:11434; the API key is optional.

Go to LLMs > Connections, pick the provider and fill in its fields, usually an API key. Credentials are stored encrypted with your ENCRYPTION_KEY.

New model connection form

You can have several connections for the same provider, say one per team or billing account. One of them is the default: it’s used when an agent names a model of that provider without picking a connection.

LLMs > Playground sends prompts straight to a model, without an agent. Use it to compare models before giving one to an agent.

  • Agents run on one or more of these connections, with fallback.