Skip to main content
You need Admin or Owner access to configure shared connections and credentials. Open AI connections in the console.

​
Choose the services

Choose a provider and model from the console, select or create a named credential, and save. Available settings depend on the provider. Confirm that your credential has access to the selected model, language, and voice.

​
Use workspace defaults

Saved workspace connections apply to agents that inherit them. Run a controlled test after changing a shared connection, particularly when several agents use it.

​
Override an agent’s connections

Open the agent’s settings and its model overrides. Enable the override for only the services you want to change. A service without an override continues to use the workspace default. For example, you can use a different voice for one agent while keeping the workspace LLM and transcriber. Save the agent and test before publishing.

​
Realtime connections

Select a realtime provider when you want it to handle speech input and output. You do not need a separate voice and transcriber for the realtime conversation. Keep an LLM configured for extraction and QA. For Google AI Studio, use a Gemini API key with access to the selected realtime model. For Vertex AI, provide the project, location, and service-account JSON required by the connection form. Your Google Cloud administrator should provide credentials with access to the selected model and region.

​
Custom Models

Use Custom Models to connect a compatible endpoint your organization operates or subscribes to. Enter the exact model ID and Base URL for each service you use. Voice also needs a supported voice ID. Select a saved credential, create one, or choose No authentication only when the endpoint intentionally supports it. The endpoint must support the relevant chat, speech, transcription, or embedding API. For chunked retrieval, embedding vectors must have 1,536 dimensions; reindex documents when changing the embedding model.

​
Verify a connection

Save the configuration, then test an agent that uses it. Check spoken responses, transcription, tool calls, and analysis as applicable. Saving a valid credential alone does not establish that the complete conversation behaves as intended.