Using Claude in a voice assistant: conversation, tools and runtime boundaries

Claude can supply the conversation model in a voice cascade: caller audio is transcribed, the model produces a response or requests an action, and a speech service speaks the answer. Choosing Claude for that role does not itself configure telephone routing, transcription, speech synthesis or interruption handling.

Burki's cascade runtime includes an Anthropic model adapter. Its supported configuration and admission checks determine what can run for your account. This is separate from Burki's GPT Live speech-to-speech mode; selecting an Anthropic model does not turn it into the same realtime session.

Design the full path

Evaluate the complete path with representative calls. A model's text capability does not establish end-to-end voice latency or caller experience.

Keep action authority explicit

Anthropic documents a distinction between client tools executed by your application and server tools executed by Anthropic. A model requesting a client tool is not proof that the requested business action succeeded. Your application must execute the permitted action and return the result. Claude tool-use documentation

A provider's server-tool catalog also does not mean every tool is exposed through Burki. Use only supported, connected actions in the selected assistant configuration.

Write instructions for a spoken task

A useful instruction fragment might be:

Collect the caller's service request and callback number.
Ask one short question at a time. Preserve corrected details.
Do not promise a booking or price from general knowledge.
Confirm an action only after its connected tool returns success.
If the action is unavailable, explain the limitation and offer the approved next step.

This is an illustrative prompt fragment, not a complete production assistant or API payload. Add your approved business facts, escalation destination and required fields.

Test the difficult turns

Have a caller correct a name, interrupt an answer and change the requested service. Make a connected action time out. Check whether the assistant stays concise, preserves confirmed facts and avoids inventing success.

Measure response onset at the caller, not just model time to first token. Recognition, end-of-turn decisions, synthesis and network playback all contribute. Compare models with the same speech components before attributing a difference to Claude.

Account for the full configuration

Include model input/output, separate speech services and telephone usage where applicable. Direct provider charges belong in the total when using BYO credentials. Use Burki's current prices and the selected provider account terms; historical model-rate tables are not a reliable admission quote.

For the surrounding architecture, read the phone-call API guide. The goal is a dependable business conversation, not merely a successful model request.