Language models - Bandwidth Agent Builder

Open AI connections → LLM to set the workspace default, or configure an agent-specific model override. You need a provider credential with access to the selected model.

Select the provider and model available in the console.
Select a named Credential or create one with your provider key.
Review any advanced response settings and save.
Test a short conversation, a tool call, and variable extraction before publishing.

A separate voice and transcriber are required for a conversation that does not use a realtime connection.

Anthropic Claude

Choose Anthropic, select a model, and select or create an Anthropic credential. Direct Anthropic credentials are separate from AWS or other gateway credentials. In Advanced settings, supported models may offer:

Response token limit to control the size of generated output.
Thinking mode and Reasoning effort to adjust reasoning behavior.
Enable prompt caching to reuse eligible prompt content.

Available settings depend on the selected model. Test the effect on response time and tool behavior with your own prompts. Cost information depends on available model pricing; custom model IDs may not have a known price.

Analysis models

You can use the conversation LLM for QA or select a separate model in the QA settings. Test extraction and analysis separately from spoken responses.