Voice agent

The Voice agent lets visitors talk to your website out loud. It answers from the same knowledge base as the chat, follows the same rules — only what your site says, no invented prices or promises — and saves each call as a conversation in your inbox.

What visitors will see

• A way to call. A microphone button in the chat panel's header starts a call. You can also show a Talk button of its own next to the chat button; on a site without the chat, it is always shown.
• A clear start. Before anything is recorded, the call panel says what will happen: the browser asks for the microphone, what the visitor says is transcribed, and how long a call can last. Visitors choose Hands-free (the assistant answers when they stop speaking) or Push to talk (they hold a button while they speak).
• An AI that says so. "You are talking to an AI assistant." stays visible for the whole call, and the default greeting says it out loud too.
• Live captions. What the visitor and the assistant say appears as text while they speak, with links to the pages an answer comes from.
• Control. Visitors can mute the microphone, switch between hands-free and push to talk, and end the call at any time. The call panel works with the keyboard alone: Space or Enter held on Hold to talk talks, Escape ends the call and closes the panel.
• The chat as a fallback. When a browser has no microphone or doesn't support voice calls, or the microphone is blocked, the panel says so. When a call can't start for another reason, such as the month's minutes being used up, it says voice calls aren't available right now. Either way, it offers the chat instead.

Set it up

You can also open the call panel from your own button:

<button type="button" onclick="asdk('talk')">Talk to us</button>

Minutes and your plan

Calls are counted in minutes, from the moment the voice provider sets a call up until it ends, on all sites of your workspace together. The Business plan includes 300 voice minutes a month and the Agency plan 1,000 (see plans and limits). The Voice page shows the minutes used this month and the calls on each site.

A call never runs past the minutes you have left: when they are almost used up, a new call is shorter, and when they are used up, new calls don't start until the next month — visitors who try are offered the chat instead. ASDK ends every call itself when its time is up. If a visitor's browser stops responding without ending the call, ASDK ends it within about two minutes, and that time counts.

How a call works

ASDK's servers set each call up with the voice provider, OpenAI (its Realtime API, over WebRTC): the widget sends its connection details to ASDK, and ASDK opens the call with your site's settings and passes the provider's answer back. The visitor's browser never gets a key to the provider. Once the call is set up, the audio goes between the browser and the provider. The provider turns speech into text and answers out loud. Whenever the assistant needs facts, it asks ASDK to search your knowledge base and answers from what comes back. ASDK keeps the transcript of the call — what the visitor said and what the assistant answered, with the pages it used — not the audio.

If your site has a Content Security Policy, Voice needs nothing beyond the widget's own entries: the browser sends its requests only to ASDK's API (connect-src https://api.asdk.com), and the call's audio is a WebRTC connection to the provider.

In your inbox

Every call appears in the inbox as a conversation marked Voice, with the visitor's and the assistant's turns as messages and the pages each answer used. You review, search and follow up on it like a chat.