Add or manage a connection
Open Providers to add a connection using the fields shown for that provider. Keep secrets in the connection form or credential manager. Do not paste API keys into prompts, tool arguments, or a modelsettings object.
An archived connection is no longer offered for new selections. Restore it if it should be available again. Before archiving a connection, check model configurations and agents that refer to it. Availability and required fields depend on the options enabled for your organization.
Create a reusable configuration
Open Models and create a configuration from the available choices. The dashboard may offer a pipeline mode, which combines language, speech-recognition, and speech-synthesis selections, or a realtime mode, which selects a realtime conversation option. The catalog shown in your organization is the source of available choices; an option documented in a general schema is not necessarily enabled for every account. Bind the configuration to the intended agent and check any workflow-level selection or supported call override. A configured default can be inherited when the workflow does not choose a more specific option. Avoid changing defaults during an active rollout unless you intend future runs to use the new selection. The workflow’s Model Configuration override is saved with the agent draft. Dashboard text says draft tests use the changed setting immediately; publish the definition for live calls to use the new workflow setting. This is separate from editing a reusable named configuration, which affects future runs that select it.Voice and language
Voice and start language are separate settings. Choose the voice for the delivery you want, then set the language the agent should begin with. Test names, numbers, interruptions, and the language transitions that matter to your callers. A catalog entry or a successful configuration save does not verify pronunciation or latency on your call path. Some organizations may see preview or account-limited options. Use the choices visible in your dashboard and verify the selected option with a controlled conversation before relying on it.LLM fallbacks
An eligible model configuration can define a backup LLM for two conditions: a Slow response that has not produced answer text or a tool call within the configured wait, or a Request error before output. The wait is configurable from 100 to 10,000 milliseconds. The primary request continues while Athena starts the fallback, and the first request to produce output serves that response. Fallbacks apply before answer output or tool execution begins; they do not switch the realtime audio model in a realtime configuration. A backup can add cost and send conversation data to the provider connection and region you configure. Changes apply to future runs. Test fallback behavior and provider access with your own account before depending on it.Updating a configuration
An edit to a reusable configuration affects future runs that select it. An already-created run retains the configuration snapshot used when it started. ForModelConfigurationOverride, use embeddings: null only when you intend to clear the optional embeddings service; explicit null for other service, mode, or configuration-reference fields is rejected. Omit fields you intend to leave unchanged.
ProviderConnectionUpdate also rejects top-level fields set to null; omit unchanged fields. Within the nested credentials object, an omitted key keeps its stored value, while a key explicitly set to null removes that stored credential. Treat credential removal as a deliberate change and never log credential values.
If a run uses an unexpected model, inspect the run configuration and check the agent selection, organization default, and any call-level overrides in that order. For call requests, see Trigger a phone call.