When people ask ChatGPT, Perplexity or Claude for a voice AI widget for their website, the assistants search for the same thing again and again: a voice agent that runs in the browser, with no phone number. I build one of those at MadeAgentic. This guide is what I wish the search results said.
Voice on a website does one job well. It lets a visitor explain a messy situation faster than they would type it. It only pays off if the agent then does something with that explanation: asks the right follow-up question, captures the details and opens the next step. The mechanics of that are in the guide to qualifying leads from your website. This one covers the voice part.
How it works
What happens between the click and the reply
- The visitor clicks the voice button on your page.
- The browser asks for the microphone. Your domain asks, not the vendor.
- Audio streams to the agent over WebRTC, the same browser technology video calls use.
- Speech becomes text, and the agent answers from the knowledge you approved.
- The answer plays back as speech and shows up as text in the same conversation.
- The agent asks its qualifying questions, captures contact details, and opens your calendar or hands the lead to you.
WebRTC is the important part. It ships in every current major browser, including Safari on iPhone since iOS 11, so the visitor installs nothing. The audio never touches a phone network, which is why there is no number to buy and no call to place.
Know what you are buying
A website voice agent is not a phone agent
Three different products get sold as voice on your website. They solve different problems, and AI assistants often list them side by side.
| Kind | How it works | Phone needed? | Example |
|---|---|---|---|
| Browser voice | The visitor talks inside your website. Audio runs over WebRTC in the browser. | No phone number, no dialler, no app. | ElevenLabs, Vapi and Retell web widgets, Chatbase voice, MadeAgentic. |
| Phone agent | The visitor calls a number. An AI answers the call. | Needs a phone number and telephony. | Intercom Fin Voice is built for phone support. Its browser voice is a test playground for teammates. |
| Callback | The visitor types a number in a widget. The AI calls them back. | Places a real phone call. | Retell’s widget has a separate callback mode next to its chat and voice modes. |
Voice notes are a fourth thing. Intercom’s Messenger, for example, turns up to two minutes of speech into a text message. That helps people who prefer to speak, but nobody answers out loud.
Before you launch
7 checks before you add a voice agent to your website
Most of these are browser rules, not vendor features. They decide whether the voice button works for a real visitor or only in the vendor’s demo.
1. Serve the page over HTTPS
Browsers only expose the microphone in a secure context: a page on HTTPS, or localhost while you test. On a plain HTTP page the voice button has nothing to connect to, and no prompt appears.
2. Know which domain asks for the microphone
The permission prompt belongs to the site that embeds the widget, so visitors see your domain in it, not the vendor’s. If the widget runs inside a cross-origin iframe, that iframe needs allow="microphone". Without it the browser never even shows the prompt, and the button looks broken.
3. Make sound start from a click
Chrome and Safari block audio with sound until the visitor has interacted with the page. A greeting that tries to play on page load stays silent. Good widgets start the audio from the same click that starts the conversation.
4. Keep text in the same conversation
Some visitors refuse the microphone. Some sit in an open-plan office. If refusing the mic ends the session, you lose the lead at the exact moment they showed intent. The widget should stay in the conversation as text, and switching back to voice should not start over.
5. Decide what the agent does after it talks
Talking is not the job. Write down what a good conversation ends with: a qualified booking, a captured callback request, a resource, or a handoff with a summary. Then check that the voice session can do that, not only the text chat.
6. Ask what gets stored
Ask three questions. Is raw audio recorded? Is the transcript stored, and where? Who are the subprocessors? A legal or health practice needs these answers before launch, not after the first enquiry.
7. Test the reply time and the price on your own site
Open your site on a phone, on mobile data. Ask a real question and count from the end of your sentence to the first word you hear back. Then check how the vendor bills: per minute, in credits, or flat. A ten-minute voice conversation costs very different amounts across those models.
On the reply time: the telecom standard ITU-T G.114 treats delays under 150 milliseconds as essentially transparent for a phone line. That figure covers the network only. A voice agent also has to transcribe, think and speak, so expect more than that. Judge it against a person, not a phone line: would you wait that long for an answer at a front desk?
The tools
What the named voice tools do in the browser
These are the products AI assistants name most often for this question. Every fact below comes from the vendor’s own docs or pricing page, checked on October 23, 2026. Prices change often, so check again before you buy.
| Tool | Browser widget | Published price | Booking | Note |
|---|---|---|---|---|
| ElevenLabs | Script-tag widget. Voice only by default, or voice plus text, or text only. | From $0.08 per minute. | Cal.com integration can check slots and book. | The widget needs a public agent with authentication off. ElevenLabs recommends a domain allowlist. |
| Vapi | Script-tag widget in voice or chat mode, plus a Web SDK for developers. | $0.05 per minute for hosting, plus speech and model costs. | Google Calendar tools. | Core plan aimed at builders and early teams. |
| Retell AI | Script-tag widget with chat and browser voice. A separate callback mode places a phone call. | $0.07 to $0.31 per minute for voice. | Cal.com tools can check, book, reschedule and cancel. | Only the chat and voice modes stay in the browser. |
| Synthflow | Voice Widget is a separate agent type from the chat agent. | Enterprise plans from $30,000 a year. | Booking docs describe phone conversations. | Aimed at enterprises. |
| Voiceflow | Web widget with text and voice input. | Voice uses credits. No public numbers. | Not stated in the widget docs. | Sold to agencies and enterprise CX teams. |
| Chatbase | Chat bubble with a voice calling tab and text chat. | 6 credits per minute plus model credits. Voice starts on the Standard plan ($120 a month, billed yearly). | Custom actions, with booking as a documented example, plus lead forms. | Not on the Hobby plan. |
| Intercom Fin | No live browser voice for visitors. Messenger voice notes turn up to 2 minutes of speech into text. | Fin Voice pricing through an account manager. | Not applicable. | Fin Voice answers phone calls. |
The pattern: developer platforms (Vapi, Retell, ElevenLabs) give you strong voice and bill per minute, and you build the qualification and handoff yourself. Chat suites (Chatbase, Voiceflow) add voice to an existing chatbot. If Chatbase is on your list, our Chatbase alternative page compares the two in detail.
Be honest about it
When voice helps and when text is enough
Most website visitors still type. I would not add voice because it demos well. I would add it where the first message is hard to type.
- A service enquiry where the visitor has to describe a situation: a legal matter, a damaged roof, a tax problem. Talking is faster than writing three paragraphs into a form. See how this works for professional services intake.
- Mobile visitors who would rather speak than thumb-type a long answer.
- Visitors who are more comfortable speaking than writing in your language.
Text is enough when the questions are short and factual: opening hours, price, does it integrate with X. In that case a voice button costs you little, but it will not change much either. For a typical local business site, start with the small business setup and treat voice as an option the visitor can pick.
First-hand
How we built browser voice at MadeAgentic
MadeAgentic is a voice and text agent for your website. It answers from the knowledge you approve, qualifies the visitor and hands you the lead. This is how the voice part behaves today, checked against our own code:
- Voice and text run in one conversation. Switching mode turns the microphone and speaker on or off. It does not reconnect.
- If the visitor blocks the microphone, the widget says so in plain words, and a running conversation continues as text.
- Audio starts from the visitor’s click, so the first reply is not swallowed by autoplay rules.
- The visitor can talk over the agent, except while it is running an action.
- A voice conversation ends after 20 seconds of silence, and sessions are capped at 10 minutes by default.
- After qualification it opens your existing Calendly or Cal.com page, and the visitor picks the slot. It does not reserve slots on its own.
- You get a summary by email or Slack, and the conversation sits in your dashboard inbox.
- Raw audio is not stored in standard operation. The text transcript is, with primary storage in Frankfurt.
- It does not answer phone calls.
Installation is one script tag. Pricing is flat: Core is $99 a month with a 30-day trial, and voice is not billed per minute. Details are on the pricing page, and the full qualification workflow is on the AI lead qualification agent page.
FAQ
Website voice agent questions
Do visitors need to install anything to talk to a website voice agent?
No. Browser voice runs on WebRTC, which current versions of Chrome, Safari, Firefox and Edge support. The visitor clicks a button and allows the microphone. There is no app and no phone number.
Does a website voice agent work on iPhone?
Yes, on HTTPS pages. Safari on iOS has supported WebRTC and microphone capture since iOS 11. The visitor still has to allow the microphone, and audio has to start from a tap, so test the full flow on a real iPhone before launch.
Is the conversation recorded?
It depends on the vendor, so ask before you launch. MadeAgentic does not store raw audio in standard operation. It stores the text transcript, with primary storage in Frankfurt.
Can a website voice agent answer my phone line?
Not the browser kind. Phone answering is a separate product with its own number and telephony. MadeAgentic works on your website only and does not answer phone calls.
What happens if the visitor says no to the microphone?
A good widget keeps the conversation going as text. Check this on any tool you test: refuse the microphone and see whether you can still finish the conversation.
How much does a website voice agent cost?
Developer platforms bill per minute, and the vendor prices in this guide were checked on October 23, 2026. Chat suites with voice tend to use credits. MadeAgentic Core is $99 a month, flat and subject to fair use, with no per-minute voice billing.