Skip to content

Configure model access

You can also describe your goal on the start page, review a plan, and generate an editable workflow. See Build workflows through conversation. Plan approval and execution confirmation are separate steps.

The model selector includes built-in Provider models and, after Team sign-in, the Server model catalog. Homepage assistants, Agent nodes, and CodeAct Agent nodes share this access mechanism without repeating model IDs in every workflow.

  1. Open Settings → Models.
  2. Enter an API key on the row for the provider you want to use.
  3. Settings save automatically.
  4. Create or edit an Agent, then choose a built-in model for that provider under Model configuration.

Ollama uses an endpoint instead of an API key. The local default endpoint is normally sufficient; enter the applicable address for a remote deployment.

  1. Connect to your team Server and sign in through Settings → Workspace.
  2. Select a model alias from the Server source on the homepage or in a node. The Server discovers the catalog from LiteLLM.
  3. Choose Chat Completions or Responses and a reported reasoning effort when the model supports them. Default reasoning does not send an extra effort parameter.

The desktop uses its current verified Workrun session; do not enter a LiteLLM Key there. Administrators configure a Team Virtual Key on Server, while LiteLLM holds Provider credentials. One Server currently maps to one LiteLLM Team, whose model permissions are shared by valid signed-in users.

Tool, image, structured-output, protocol, and reasoning-effort capabilities come from gateway metadata; unknown capabilities are not guessed. Model aliases must match the Team Key’s visible catalog. Responses uses client-side history replay, with store: false enforced by Server.

Team mode still permits explicitly selected local Provider connections. Direct requests use local credentials and bypass Server/LiteLLM; gateway failures do not automatically switch to direct access. Switching accounts or signing out interrupts team-model streams. Background recovery also requires the original user’s valid session; a persisted task does not authorize another account.

See Deploy Server, LiteLLM, and Langfuse. Old workflows do not automatically map personal models to team aliases: select the intended model, save, and publish a new version when needed.

  • One provider key can be used by that provider’s built-in models.
  • Never put keys in node instructions, workflow inputs, or Python code.
  • Configure an Agent’s name, role, instruction, tools, and generation parameters in the node inspector.
  • After switching models, rerun the workflow with the same inputs to confirm the result.

Model configuration in the Agent inspector, showing the built-in catalog grouped by provider.