---
title: "Voice calls for Chat Thing AI agents on your website"
canonical_url: "https://chatthing.ai/blog/introducing-voice-calls"
last_updated: "2026-10-21T15:34:32.407Z"
meta:
description: "Let website visitors talk to your Chat Thing agent. It answers from your knowledge, runs your power-ups and keeps your choice of AI model."
"og:description": "Let website visitors talk to your Chat Thing agent. It answers from your knowledge, runs your power-ups and keeps your choice of AI model."
"og:title": "Introducing voice calls: your AI agent, now speaking - Blog - Chat Thing"
"twitter:description": "Let website visitors talk to your Chat Thing agent. It answers from your knowledge, runs your power-ups and keeps your choice of AI model."
"twitter:title": "Introducing voice calls: your AI agent, now speaking - Blog - Chat Thing"
---
[← Back to blog](https://chatthing.ai/blog)**Blog**# **Introducing voice calls: your AI agent, now speaking**
Zef
21 Oct 2026
~ 9 min read
---
---
### **On this page**
1. [A conversation inside your website](#a-conversation-inside-your-website)
2. [Watch it search for a home, then open the map](#watch-it-search-for-a-home-then-open-the-map)
3. [From a question to an enquiry](#from-a-question-to-an-enquiry)
4. [How it works: voice and reasoning have different jobs](#how-it-works-voice-and-reasoning-have-different-jobs)
1. [You keep your choice of model](#you-keep-your-choice-of-model)
2. [Calls keep moving while work happens](#calls-keep-moving-while-work-happens)
5. [Your website can join the conversation](#your-website-can-join-the-conversation)
6. [Make it sound like your agent](#make-it-sound-like-your-agent)
7. Where we'd put it to work
1. [Qualify sales enquiries while interest is high](#qualify-sales-enquiries-while-interest-is-high)
2. [Talk someone through a support problem](#talk-someone-through-a-support-problem)
3. [Help shoppers narrow things down](#help-shoppers-narrow-things-down)
4. [Turn service questions into enquiries](#turn-service-questions-into-enquiries)
5. [Make a complex app easier to explore](#make-a-complex-app-easier-to-explore)
8. [What it doesn't do yet](#what-it-doesnt-do-yet)
9. [Pricing](#pricing)
10. [Try it](#try-it)
Some requests are easier to say than to type.
You're looking for a home near the sea, with at least two bedrooms, within your budget. Then you spot somewhere interesting and want to know more. Where is it on the map? What's nearby? Can you save it for later?
With voice calls, visitors can have that conversation with your Chat Thing agent while your website responds around them. They ask a question, add a detail, change their mind. Your agent answers from its knowledge, uses its power-ups and talks them through the result, with the same AI model you've already chosen doing the thinking.
## A conversation inside your website
Voice calls add a call button to your agent's web chat. Visitors allow microphone access, start talking and hear the agent reply in the same window. The call has a live visualisation in your agent's colours, listening and speaking states, a mute button, optional captions and a transcript download.
People can interrupt to clarify what they meant, ask a follow-up without repeating the whole request, then end the call and carry on typing in the same conversation.
That continuity matters. If someone has already explained their problem in text, starting a call shouldn't mean starting again. The call picks up the conversation that's already on screen, and everything said is saved back to the chat.
_The call interface in our Cornish Horizons demo. Colours follow the agent's branding._
## Watch it search for a home, then open the map
Our Cornish Property Dreams demo has a property catalogue, filters, a map and a shortlist, and its agent already has website power-ups for all of them. A call gives visitors another way to use them.
In the recording, we ask for homes with at least two bedrooms and sea views, talk through a few, then ask to see them on a map. The agent acknowledges the request, the page switches from property cards to map markers, and the conversation carries on beside it.
_A continuous 14-second excerpt at normal speed, including the wait. The properties are demo data._
"Show me some on a map" is enough because the agent acts on the results already in the conversation. We also ask about one particular property, and the agent talks through the balcony, kitchen and parking without leaving the call.
For a visitor, that means explaining what matters to them while looking at the actual options. The page shows the results and the conversation helps them explore.
## From a question to an enquiry
Our second demo, Willow Studio, is a fictional photography and podcast studio. Its agent has a studio handbook as a data source, with room prices, equipment and policies, so a visitor can ask what a session costs, what's included or which room suits them.
From the handbook, the agent works out that two hours in the Daylight Studio plus the paper backdrop pack comes to £111. It can then open and fill the booking form on the page, where the visitor can check the room, duration and details, and send the enquiry on through an API power-up.
That gives the conversation somewhere useful to go. A visitor can get from "would this work for me?" to a filled enquiry form without hunting for the right page and starting again.
Willow's enquiry endpoint is a test one, so nothing is really booked and no payment is taken. A live booking service would bring its own availability checks and booking rules.
## How it works: voice and reasoning have different jobs
Voice calls run on two layers, and that's the part of the design we're most pleased with.
OpenAI's GPT-Live-1 handles the spoken conversation: listening, speaking, acknowledging requests and coping with interruptions. When a question needs your business information, or someone asks for something to be done, it hands the request to your Chat Thing agent running behind it. That handoff is called delegation, and OpenAI [designed GPT-Live around it](https://developers.openai.com/api/docs/guides/live).
Your agent then does what it always does, with its own model, instructions, knowledge and power-ups, and passes the result back for the voice to say.
_Our explainer shows the two layers. GPT-Live-1 is the voice; the agent uses a model supported by Chat Thing and available on your plan._
Here's the map request, step by step:
1. You speak. Your browser carries the audio in real time using WebRTC, the same technology browsers use for video calls.
2. The voice layer delegates the request, along with the conversation so far and any page context your integration is allowed to share.
3. Your agent picks the right tool, in this case a website power-up that changes the view.
4. The website runs the action and reports back. The map opens in the page you're already on.
5. The voice tells you what happened. It's instructed not to claim an action worked until the agent confirms it.
A policy question uses your knowledge in the middle instead, and an enquiry calls your API. The voice stays the same while the work behind it changes.
### You keep your choice of model
Adding voice doesn't move your agent's thinking to the voice provider. If your agent runs on a supported Claude, Gemini or OpenAI model, that model keeps doing the reasoning, and GPT-Live-1 is the voice around it.
So you can choose the model for the job and the voice for the experience. An agent that needs careful answers from a detailed knowledge base may want a different model from one that mostly helps people browse a catalogue, and neither has to change because you've added voice.
### Calls keep moving while work happens
Looking things up takes time. The voice can acknowledge a request while your agent works, and Chat Thing passes answer sentences back as they're ready rather than waiting for the whole answer.
The two layers also get different instructions. The voice instructions shape personality, tone and speaking style, while your agent keeps the detailed business instructions and tool workflows. The spoken side can stay short and friendly without the substance drifting away from your business.
## Your website can join the conversation
The property demo works because its website exposes deliberate actions through the Chat Thing SDK: search the catalogue, change preferences, show results, open a map. Voice calls use those same actions.
If you've already built website power-ups for your agent, voice gets them for free. If you haven't, you decide what the agent can operate. A shopping assistant might filter products and open a comparison. A reporting tool might change a date range and redraw a chart. A service business might have it fill in an enquiry form.
Your website still has to implement each action and allow it in the SDK. Voice gives people a conversational way into them.
For anything that needs a confirmation, build that into the action or the workflow. Interrupting the agent is useful for correcting a conversation, but it can't recall an API request that has already been sent. Voice follows the same power-up rules as text chat.
## Make it sound like your agent
Voice settings live in your agent's web channel:
- 13 voices with British, Irish, Australian, North American and other accents, with previews that don't use minutes.
- Voice instructions for personality and tone, drafted from your agent's system prompt and then yours to edit.
- Spoken language: your web chat's language, a fixed language, or follow the visitor.
- Answer length: short (up to 90 words), medium (180) or long (300).
- A maximum call length of 1 to 15 minutes. Visitors see a countdown, and the agent is asked to wrap up before time runs out.
- An optional monthly limit per agent, so one agent can't use the whole team allowance.
_Settings from our Willow Studio test setup. The limits shown are test values._
A support agent might give one short instruction at a time. A product adviser might lay out a few options and ask what matters most. The settings let you shape how each of those conversations sounds.
## Where we'd put it to work
### Qualify sales enquiries while interest is high
"We're looking for something for our support team. Would this be a good fit?"
A visitor on your pricing page may be ready to talk right now. With your qualification criteria in its instructions, an agent can ask about their use case, team size and timeline while answering product questions, then capture the details in your CRM or open a booking link through the integrations you've connected. Your sales team gets the context before the first meeting.
### Talk someone through a support problem
"I'm on the settings page, but I can't find where to change this."
Your agent reads your help content and talks through the next step. With the right website power-ups, it can open the panel itself. People can describe what they see instead of typing it out between every step.
### Help shoppers narrow things down
"I need something for a small desk, under £150, ideally in black."
Connect catalogue search and filters, and let the conversation refine the results while people look at the products.
In our Skin Routine Finder demo, a shopper describes their skin type, concerns and budget, gets a suggested routine, then changes it three times without starting again: checking the ingredients, swapping out a brand and asking for a cheaper SPF. When the catalogue can't confirm something, such as whether every product is fragrance-free, the agent says so and suggests checking the pack.
_One continuous conversation, recorded shortly before launch. The products come from the demo's catalogue._
### Turn service questions into enquiries
"Would this package work for a team of eight?"
Give your agent your service details and an enquiry form or API. It can explain the options, clarify what the customer needs and collect the details your team needs to reply.
### Make a complex app easier to explore
"Show me last month's results, grouped by region."
Expose the relevant filters and view controls as website power-ups. The agent turns the request into changes people can see, while they keep using the normal interface.
## What it doesn't do yet
- If a teammate takes over a conversation, they carry on in text. Team members can't join a live call.
- Calls are capped at 15 minutes.
- Your agent will sometimes misunderstand a request, as people do. In our property recording it briefly took a request about saving a favourite as a question about gyms and needed a clarification. Try the conversations that matter before switching voice on for everyone.
## Pricing
Voice calls are a monthly add-on to any paid Chat Thing plan. A voice pack gives your team 60 minutes of calls a month for $10, shared across all your agents. Buy as many packs as you need.
Minutes reset each month and don't roll over. There are no overage charges: when minutes run out, new calls are politely declined and text chat keeps working. If you're running low before your renewal, a team owner can use **Add minutes now** on the billing page to buy extra packs at the full pack price, and the minutes are ready straight away. The work your agent does during a call, such as looking things up and running power-ups, uses your normal message tokens, the same as in chat.
During launch, eligible teams get $5 off their first voice invoice, so your first pack costs $5, then $10 a month. The discount applies once per team, however many packs you buy.
## Try it
Add a voice pack under **Billing → Voice calls**. Then open your agent's **Channels** tab, click the settings icon on the **Web** card and find the **Voice calls** section. Turn on **Enable voice calls**, pick a voice, check the voice instructions and click **Update settings**.
Open your web chat and try three things: a question from your knowledge base, a follow-up, and one action you've set up. Start with a conversation you already understand well, and give people a way to ask, see the result and keep going.
[Open Chat Thing](https://app.chatthing.ai/app/bots) to give your agent a voice.
Your agent can answer the phone too, with the same voice engine and settings. See [how to set up the phone channel](https://chatthing.ai/docs/channels/phone).
### **Related articles **
[](https://chatthing.ai/blog/chat-thing-webmcp-website-copilot) [Turn a WebMCP-enabled website into a copilot with Chat Thing
](https://chatthing.ai/blog/chat-thing-webmcp-website-copilot)
Chat Thing can now discover the tools a WebMCP-enabled page exposes and let an embedded agent use them. Add the widget, switch on WebMCP, and the page has a conversational copilot.
[](https://chatthing.ai/blog/supportbench) [SupportBench: we tested 7 AI models as customer support agents
](https://chatthing.ai/blog/supportbench)
We tested seven AI models on 1,085 support conversations. Gemini 3.7 Flash led the main score, while Grok 4.6 wrote the preferred reply most often.
[](https://chatthing.ai/blog/blog-coworking-space-agent) [Our co-working space down the road just got an AI agent. Here's what happened.
](https://chatthing.ai/blog/blog-coworking-space-agent)
C-Space is a coastal coworking space in Newquay with hot desks, private offices and event hire. This case study looks at how their AI agent answers the everyday questions, parking, opening hours, walk-ins, the moment they come in, so the team can stay focused on running the space instead of sitting on the inbox.
- **Previous **
- [← Inline citations, phone call summaries and eight new models](https://chatthing.ai/blog/citations-call-summaries-and-eight-new-models)