Build an OpenAI voice agent for your website

To build an OpenAI voice agent with audentic, create a private draft from a public page on your website, review its business knowledge, test a conversation, and publish it to the website addresses you approve. Then copy the website widget or add the React component. audentic manages the OpenAI connection and keeps model credentials on its server.

audentic is an independent platform built around OpenAI voice models. It is not an OpenAI product. This guide uses the Realtime conversation option for a website agent that can answer questions and help visitors send a follow-up request.

Create a private voice agent, or try an example in the gallery before building your own.

Choose a managed builder or a custom API integration

Use audentic when you want to run a website agent without building the voice infrastructure. The workspace brings together instructions, reviewed business knowledge, a private preview, website permissions, the widget, and saved conversation review. You configure and publish the agent in audentic; a coding assistant can install the published agent on your website.

Build directly with the OpenAI API when you need to own the application and its behavior. You implement session creation, the audio interface, business tools, authorization, and operation of the service. OpenAI documents a browser path using its Agents SDK and WebRTC, as well as lower-level connections for more control. Standard API keys belong on your trusted server, never in browser code or a public repository. See OpenAI's Realtime getting-started guide and WebRTC connection guide.

The Realtime API supports direct audio conversations and tool calls. Choosing a model does not supply your business facts or make actions such as bookings work automatically; those require application logic and an authorized integration. audentic's current scope is described below.

1. Start with one useful job and a website page

Open Create an agent, enter a public page address, and select Create private draft. Use a page that explains the service visitors ask about, such as a clinic's services or a gym's membership page.

The import reads that page. It does not crawl your whole website, sign in to private pages, or run the page's JavaScript. Check the source summary and any warnings. A private draft is available for your own testing; creating it does not publish a widget to visitors.

If you do not have a suitable page, choose Use voice or type instead. Describe who the agent helps and its job in plain language. For example: “Help visitors understand our classes and send our team a request for a tour.” You can complete the setup by typing without microphone access.

2. Review the facts and the starting language

In About, give the agent a clear purpose and set its Starting language. Put detailed behavior instructions in Extra guidance and business facts in Knowledge. For example, tell it to ask which service a visitor needs, use the published opening hours, and offer a human follow-up when information is missing.

Check imported prices, opening hours, addresses, and contact information against your current business information. Add missing facts before testing. The agent is instructed to adapt to the visitor's preferred language, but you should test the languages and names your visitors use.

In Help & voice, choose a voice and hear a sample. This guide focuses on the Realtime model options. The configurable GPT-Live option has a different architecture; it should not be treated as the same Realtime configuration. See agent configuration and knowledge setup for the available controls.

3. Test privately before publishing

Open Test & publish and try the suggested questions, then test these situations:

  • A question answered by your knowledge, such as “What time do you close?”
  • A missing fact, such as a price you have not provided. Check that the agent acknowledges the gap.
  • A request outside its job, such as changing an existing booking without a booking integration.
  • Your starting language, a visitor switching languages, and a spoken interruption.
  • A follow-up request, if enabled. Check what the visitor reviews and sends, then inspect the saved request.

Choose Type instead to test answers without granting microphone access. Typed conversations still connect to the model and use a conversation start. Written practice runs can help check instructions and facts; you also need a spoken conversation to check voice delivery, interruptions, and microphone behavior.

End the active preview before editing, correct the instructions or knowledge, and try the same questions again. Review conversations and requests to understand what is saved.

4. Publish and add the website widget

In Website addresses, approve every exact website origin that will host the agent. For example, https://example.com and https://www.example.com are separate entries. Add a staging origin separately if you use one. Save your changes, then select Publish agent.

Select Copy snippet and paste the generated code once into your website's custom-code or footer-script area. It includes your agent ID and the correct audentic host. Your website must allow scripts. For a React app, install @audentic/react and use the published agent ID in its SessionControl component. No OpenAI API key belongs in either integration.

Follow the website widget guide for the snippet and supported placements, or the React integration guide for component code. If an assistant is editing your site, use the coding-agent instructions and downloads. Guides are also available for Lovable, v0, Bolt, and Replit.

5. Try the live page as a visitor

After deploying your website, visit the published page while signed out of audentic. Visitors do not need an audentic account. Opening the panel does not start the microphone; the visitor chooses to start a conversation and grants access. Try speaking, interrupting, ending the session, and Type instead. Test on a mobile device too.

Voice requires HTTPS or localhost, a compatible browser, and microphone permission. An iframe or a restrictive website security policy can require additional configuration. The builder's Check installation can find the snippet in public page HTML; it does not verify a working microphone or conversation. Use the browser troubleshooting steps if the workspace preview works but the live page does not.

Frequently asked questions

Do I need my own OpenAI API key?

No. audentic currently provides the model access for its managed website agents and does not accept customer API keys. Current availability and usage limits are covered in usage and plans. A custom application built directly on OpenAI's API needs its own server-side API access.

Can visitors use the agent without speaking?

Yes. Visitors can choose Type instead without granting microphone permission. A working internet and model connection is still required. Typed testing helps check answers, but it does not establish how the agent sounds or handles spoken interruptions.

Can the agent speak another language?

Set Starting language in the agent's About settings. The agent is instructed to open in that language and adapt to a visitor's preference. Test the languages, pronunciation, and business terminology you need before publishing; a language setting is not a guarantee of correct answers.

Does importing my website connect bookings, payments, or customer accounts?

No. Importing a page supplies reference information. It does not authorize the agent to change external systems. audentic can support configured on-screen interactions and visitor-confirmed follow-up requests; bookings, payments, and account changes need a supported integration. Website agents are the current focus, rather than outbound telephone campaigns or a complete CRM.

Can I use a coding assistant to create and publish the agent?

Create, review, and publish the agent in the audentic workspace. A coding assistant can then install an existing published agent on your website using its ID, the widget, or the React component. The coding-agent guide includes reusable integration instructions. A supported public management API, CLI, and MCP server are planned; they are not available as a substitute for the workspace workflow today.

How do I make the agent more useful after launch?

Review saved conversations for unanswered questions, incorrect facts, and requests that need a human. Update the knowledge or guidance, repeat the relevant private tests, and verify the change on your live page. Keep its first job narrow enough that you can tell whether it helped a visitor.

Create your private draft and test the first question your visitors usually ask.

Updated · Written by the audentic team