What happens during a conversation?

A visitor clicks the button on your site and types what they need. The widget tells them they’re speaking with AI, that a transcript will be saved and how long the call can last. Once they allow the microphone, their browser streams the audio to OpenAI’s voice service over WebRTC, and the agent greets them in English.

As they talk, the agent answers aloud, and it’s told to search the approved knowledge for its channel whenever a question needs product facts. The visitor can switch on live captions. If nothing in the knowledge answers the question, it’s told to say so and suggest your usual contact options, or offer to connect the visitor to your team if you’ve turned on human handover, an early release. The call ends when they hang up or at the 10-minute limit, and the transcript appears in your dashboard.

1. Add and approve your knowledge.

Once James has opened your workspace, add the material the agent may use: FAQs and notes you write, or selected .md, .mdx and .txt files from GitHub for support. Sales gets its own notes on features, plans and prices.

Nothing is used until it’s approved. Adding a source never approves it, whether it comes from the dashboard, MCP or the REST API, and every GitHub re-sync needs approval again.

2. Configure your voice agent.

Set the agent’s name, greeting, instructions and voice, then the button label and colour. Support and sales are configured separately, so each can greet visitors and answer in its own way.

3. Add your website and the embed code.

List the exact origins where the widget may run, such as https://www.example.com, up to 10 for each agent. Copy the script tag from the dashboard and paste it just before the closing </body> tag. The snippet holds only your public workspace ID, never a secret key.

4. Test a support call.

Start a test call from the dashboard. Ask the questions your customers ask, in different ways, then read the transcript and fix any gaps in the knowledge. Tests are marked as tests and count towards your 50 free conversations.

5. Switch on customer calls.

Customer calls start switched off. Turn them on when the answers are ready, and calls can then start from the origins you listed. The checklist on your dashboard Overview tracks these five steps for the support agent.

Review conversations and keep knowledge current.

Read transcripts in the Conversations tab, filtered by support or sales, and check the Insights tab for recurring questions and reported issues from recent customer calls. When the agent couldn’t answer, add a source with the missing answer and approve it, and remove any source it replaces. Re-sync GitHub docs after they change, and approve the update before the agent uses it.

You can also do this from your own tools. A remote MCP server and a REST API let a coding agent add knowledge, change settings, and read conversations and automatic insights with a workspace token.

Common questions

What does a visitor see before the call starts?

The Kawl Now widget asks what they’re calling about and shows this notice under the start button: “By starting, you agree to microphone use and a saved transcript. You’re speaking with AI. Up to 10 minutes.” When they start, the browser asks for microphone access.

Which setup steps can a coding agent do?

Most of them. With a workspace token, a coding agent can use Kawl Now’s MCP server or REST API to add knowledge, update agent settings and allowed origins, switch calls on and read conversations, and the MCP get_installation tool returns your embed code. Approval is always a separate call, but any workspace token can make it. If a person should approve each source, block the approve_knowledge tool in your MCP client and tell your coding agent to leave approval to you.

Will joining the waitlist install anything on my site?

No. Joining the waitlist only sends us your details, and we open your Kawl Now workspace with you. You choose the knowledge, list the allowed website origins and add the script tag yourself, and customer calls stay off until you switch them on.

Should I test before switching on customer calls?

Yes. Customer calls stay off until you switch them on, and they only start from origins you’ve listed. The Kawl Now dashboard checklist walks you through approving knowledge and a test call first. The REST API and MCP don’t check those steps, so make your test call before you turn calls on.