How do the three routes compare?
Each route can put a voice agent on your website. They differ in how much of the surrounding system your team builds and runs. Prices are as listed on 8 October 2026.
| Consideration | Build it yourself | Voice agent platform | Kawl Now |
|---|---|---|---|
| Your team’s work | Write and host the session server, interface, retrieval, review tools and limits. | Configure the prompt, knowledge, tools and channels in a dashboard or by API. | Approve knowledge, add your allowed sites and test, then switch on customer calls. |
| Website widget | Your own, or LiveKit’s hosted embed for agents on LiveKit Cloud. | A hosted widget with voice and text options, added with a code snippet. | One script tag per agent, with browser voice and live captions. |
| Which sites can start a call | Origin checks, authentication and request limits that you add. LiveKit Cloud’s embed widget has allowed origins built in. | Domain allowlists, plus optional reCAPTCHA at Retell AI. Voiceflow’s allowlist needs a Business or Enterprise plan. | Up to 10 exact HTTPS origins per agent, plus one-use call tickets. |
| Knowledge | Any retrieval you choose, connect and host. | Built-in knowledge bases. Retell AI, ElevenAgents and Voiceflow can refresh sources on a schedule. | Manual notes, plus selected GitHub docs for support, each approved before use. |
| Channels | Browser, app or phone, as you build it. | Web and phone, plus SMS, WhatsApp or email on some. | Browser voice on your website only. |
| Tools and integrations | Anything your code can call, including account data. | Tool calls plus CRM, help desk or calendar integrations, varying by vendor. | No custom tools, CRM, calendar or account lookup. |
| Cost to start | Model usage plus hosting. OpenAI lists GPT-Live at $0.05 a minute plus backend usage, and prices Realtime API models such as gpt-realtime-2.1 per token. LiveKit’s free Build plan includes 1,000 agent session minutes a month, with $2.50 of LiveKit Inference credits covering about 50 minutes of model use. | Free credits or minutes first: $5 at Vapi, $10 at Retell AI and 15 minutes on ElevenAgents’ free plan. Voiceflow offers agencies a free trial and asks businesses to book a demo. | First 50 conversations free per workspace, tests included. Paid calls aren’t available yet, so new calls stop after that. |
- OpenAI API pricing
- LiveKit pricing
- LiveKit free plan allowances
- LiveKit Agent Embed Widget
- Vapi web widget
- Vapi API keys
- Vapi knowledge bases
- Vapi phone calling
- Vapi integrations
- Vapi pricing
- Retell AI homepage
- Retell AI website widget
- Retell AI public keys
- Retell AI knowledge base
- Retell AI pricing
- ElevenAgents
- ElevenAgents widget
- ElevenAgents knowledge base
- ElevenAgents scheduled URL sync
- ElevenAgents pricing
- Voiceflow widget install
- Voiceflow widget settings
- Voiceflow knowledge sources
- Voiceflow phone numbers
- Voiceflow pricing
How do you build an AI voice agent for your website?
OpenAI’s GPT-Live and Realtime API connect a browser to a voice model over WebRTC. Your server keeps the API key and either exchanges the browser’s connection offer or creates an ephemeral key. OpenAI’s WebRTC guide includes example server and browser code, and its Agents SDK gives the Realtime API a browser-first starting point. LiveKit Agents and Pipecat are open-source frameworks that run your voice agent code with the providers you choose.
LiveKit Cloud adds a hosted embed widget with allowed origins, and its observability includes transcripts. Pipecat’s docs show how to save transcripts and list a community-maintained retrieval service you can add to your pipeline. OpenAI’s guide says to protect your session endpoint with authentication, authorisation, request limits and HTTPS before other people use it. Whichever you pick, your team builds or assembles these parts and keeps them running:
- Short-lived session access that keeps provider keys out of the browser.
- Origin checks, so only your sites can start calls.
- Retrieval that finds the right passage in your docs.
- A review and approval step for sources.
- Storage and a review screen for transcripts.
- Limits on call length, concurrent calls and spend.
- Abuse protection for a public microphone button.
When is a voice agent platform the better fit?
A platform suits you if you need phone numbers, text chat beside voice, a choice of model or tools that act in other systems. The vendor hosts the infrastructure, and you design and test the agent. Each of these documents a website widget:
- Vapi’s web widget adds chat and voice to a website with one line of code, and its Web SDK starts calls from your own browser interface.
- Retell AI’s web calls run in the browser with no phone number involved, and its script-tag widget offers chat, voice and callback modes.
- ElevenLabs says ElevenAgents embeds with one line of HTML. Its widget defaults to voice only, and its agents also run on phone, WhatsApp, SMS and email in more than 90 languages.
- Voiceflow’s widget defaults to chat, and its voice modality makes the conversation behave like a phone call.
What does Kawl Now include?
Kawl Now runs on OpenAI’s GPT-Live over WebRTC, so OpenAI processes the voice conversations. Around the model, Kawl Now gives you the parts a team would otherwise build:
- Separate support and sales agents, each with its own knowledge, greeting, instructions, voice and allowed sites.
- A one-line embed that carries only a public workspace ID, so provider keys stay on the server.
- Retrieval that searches only the sources your team has approved, with fresh approval needed after every GitHub sync.
- A 10-minute limit per call, with up to 3 calls at once and 12 new calls a minute per workspace.
- Text transcripts and caller feedback in the dashboard. Callers are told they’re speaking with AI, and Kawl Now doesn’t store call audio.
- A REST API and an MCP server for coding agents. Adding a source never approves it, whether through the dashboard, MCP or REST; approval is a separate step.
What does Kawl Now leave out?
If you need any of these, a platform or your own build fits better:
- A phone number for customers to call, outbound calls to customers, or text chat.
- CRM, calendar or help desk connections, custom tools or account lookup.
- A choice of model, a language setting or more than two voices.
- Website URL import. Paste website content into a sales note instead.
- Call routing with queues or business hours. Human handover is an early release that rings one verified UK or US number per agent, one handover at a time, and customer routing is still being tested.
- A CAPTCHA or other bot check before calls start.
- Calls beyond the free allowance. Paid calls aren’t available yet, so new calls stop once a workspace’s 50 free conversations are used.
Which route suits your team?
Build it yourself if voice is part of your product, the agent needs your account data or custom tools, or you want full control of the models, interface and costs.
Use a platform if you need phone calls, text and voice together, many languages or CRM and help desk integrations. Allow time to configure and test the agent.
Install Kawl Now if you want spoken answers to product and sales questions on your SaaS website, from sources your team approves. It runs the session server, call limits and transcript review for you, so your team can focus on the knowledge and the testing.
Common questions
What do you need to build an AI voice agent for a website?
A voice model or framework, such as OpenAI’s Realtime API, LiveKit Agents or Pipecat, plus a server that keeps API keys out of the browser, origin checks so only your sites can start a call, search over your docs, transcript storage, and limits on call length and spend. If you’d rather not build those parts, our guide to adding a voice agent to your website shows the ready-made route with Kawl Now.
Can you build an AI voice agent without writing code?
Yes, on a hosted platform. Retell AI says you can build entirely from its dashboard with no code, ElevenLabs says a first agent needs no engineering, and LiveKit’s Agent Builder lets you prototype agents in the browser. You still write the prompt and set up knowledge and channels. Kawl Now starts you with a support agent and a sales agent, so you approve knowledge, set a greeting and voice, add your site as an allowed origin, paste one script tag and test before you switch on customer calls.
What does it cost to build an AI voice agent?
Model usage, hosting and your team’s time. On 8 October 2026, OpenAI listed its GPT-Live model, gpt-live-1, at $0.05 a minute plus backend usage, and priced Realtime API models such as gpt-realtime-2.1 per token, at $32 per million audio input tokens and $64 per million audio output tokens. LiveKit’s free Build plan included 1,000 agent session minutes a month but only $2.50 of LiveKit Inference credits, about 50 minutes of model use, so beyond that you pay for models on a paid plan or through your own provider accounts. Vapi and Retell AI start with free credits, then charge per minute. Kawl Now gives each workspace its first 50 conversations free, including tests. Paid calls aren’t available yet, so new calls stop after that.
Can you build an AI voice agent that answers phone calls?
Yes. Pipecat supports dial-in and dial-out over phone lines and SIP, LiveKit’s free Build plan includes one US phone number, and platforms such as Vapi and Retell AI provide numbers too. Kawl Now works only in the browser on your website, so it has no phone number for customers to call and doesn’t answer phone calls.