How do Deepgram’s Voice Agent API and Kawl Now compare?
Deepgram sells the voice pipeline and the tools to build on it, while Kawl Now is a finished agent for one job. We checked Deepgram’s official docs and pricing on 8 October 2026.
| Consideration | Deepgram Voice Agent API | Kawl Now |
|---|---|---|
| What it is | One API that runs speech-to-text, a language model and text-to-speech. | Ready-made support and sales voice agents. |
| On your website | A browser SDK and a widget with six layouts, connected with tokens from your server. | One script tag, with no server of your own. |
| Models | Deepgram Flux or Nova-3 to listen, an LLM from OpenAI, Anthropic, Google and others or your own, and Deepgram or third-party voices. | OpenAI’s GPT-Live, with two voices. |
| Knowledge | Your own search, reached through function calls. | Approved FAQs, notes and GitHub docs. |
| Phone calls | Inbound and outbound, through Twilio, Genesys, Amazon Connect or AudioCodes. | No phone number for customers to call. Visitors talk in the browser. |
| Languages | Multilingual speech recognition and voices, chosen per agent. | Agents greet callers in English. |
| Transcripts | Sent as events during the call. You store them. | Saved to your dashboard. |
| Calls at once | Up to 45 on pay as you go. | Up to 3 per workspace. |
| Longest session | 2 hours. | 10 minutes. |
| Price to start | $0.075 a minute on the Standard tier, after $200 of free credit. | First 50 conversations free per workspace, a one-time allowance. Paid calls aren’t available yet. |
What is the Deepgram Voice Agent API?
It replaces the work of stitching separate speech services together. You open one WebSocket, send a Settings message that names the listen, think and speak providers and the prompt, then stream audio in and receive audio back. Deepgram handles end-of-turn detection, barge-in and function calls, and you can update the prompt or swap a model mid-conversation.
Under the hood it is a chain rather than a single speech-to-speech model. Deepgram Flux or Nova-3 turns the caller’s words into text, the language model writes a reply, and Deepgram’s Flux TTS or Aura voices, or another provider’s, speak it. Deepgram says keeping all three stages behind one API avoids network round trips between them, and you can bring your own LLM or voice provider while keeping its orchestration.
How do you add Deepgram’s voice agent to a website?
Deepgram’s Browser Agent SDK comes in four layers: a core WebSocket client, React hooks, React components and a drop-in widget. The widget loads from Deepgram’s CDN or npm with one init call, offers six layouts such as a floating button, a sidebar or an orb, and can show a transcript and a text box. You can change its labels and set its colours and fonts with theme tokens.
Your API key must never reach the browser, so the widget asks a token factory for a short-lived token before each connection. That factory is an endpoint on your own server, which you build and protect, along with any limit on who can start calls. Deepgram also notes that its dashboard shows high-level usage only, and recommends saving every non-audio WebSocket message yourself if you want transcripts.
Kawl Now’s embed holds only a public workspace ID. Calls start only from the sites you list, through one-use call tickets, and each conversation’s transcript appears in your dashboard.
Where does each agent get its answers?
Deepgram leaves knowledge to you. Within a turn, the language model can call your functions to reach databases, embedding models or search APIs. That gives you full freedom, but your team builds and hosts the search and decides how new content gets checked.
Kawl Now’s support agent searches FAQs, notes and selected GitHub .md, .mdx or .txt files, and the sales agent has its own notes. Each source waits for approval before an agent can use it, and every GitHub re-sync needs approval again.
How much does each cost?
On 8 October 2026, Deepgram’s pricing page listed the Voice Agent API on pay as you go at $0.075 a minute for the Standard tier, the $4.50 an hour its product page quotes, and $0.163 a minute for the Advanced tier. Bringing your own LLM and text-to-speech lowers the rate to $0.050 a minute, though you then pay those providers yourself. Time is measured by WebSocket connection, new accounts get $200 of free credit, and pay as you go allows up to 45 concurrent agent connections.
The tier depends on the language model: Deepgram lists models such as gpt-5.4-mini as Standard and gpt-5.4 as Advanced.
Kawl Now gives each workspace its first 50 conversations free, shared by support, sales and tests, with no card. Paid calls aren’t available yet.
When is Deepgram the better choice?
Deepgram suits developers building their own voice agent who want to choose each part of it. Look at it first if you need:
- Your own choice of speech, language and voice models, including your own LLM.
- Phone calls through Twilio, Genesys, Amazon Connect or AudioCodes, inbound or outbound.
- Conversations in more than one language.
- Dedicated, VPC or self-hosted deployment, or processing on Deepgram’s EU endpoint.
- Certifications for a security review: Deepgram lists SOC 2 Type 1 and Type 2, and signs HIPAA agreements with Enterprise customers.
When does Kawl Now make more sense?
Kawl Now suits a SaaS team that wants spoken answers to support and sales questions on its website, without running a token server, a search service and a transcript store. It has no phone number for customers to call, no model choice and no custom tools, and it greets callers in English. It gives you:
- Support and sales agents with separate knowledge, greetings, voices and widgets.
- Search limited to sources your team has approved.
- Limits built in: up to 10 allowed sites per agent, 10-minute calls and three at once.
- Transcripts in the dashboard, plus a REST API and an MCP server for coding agents.
Common questions
Does the Deepgram Voice Agent API have a website widget?
Yes. Deepgram’s @deepgram/agents-widget package adds a voice agent to a page with one init call and six layouts. It needs a token factory, an endpoint on your server that fetches short-lived tokens, because the API key must never reach the browser. Kawl Now’s widget needs no server of yours.
How much does the Deepgram Voice Agent API cost per minute?
On 8 October 2026, Deepgram listed $0.075 a minute for the Standard tier on pay as you go, $0.163 for the Advanced tier and $0.050 if you bring your own LLM and text-to-speech. Time is measured by WebSocket connection, and new accounts start with $200 of free credit.
Is the Deepgram Voice Agent API speech-to-speech?
Audio goes in and audio comes out, and Deepgram describes it as a unified speech-to-speech API. Inside, it chains speech-to-text, a language model and text-to-speech, with turn-taking and interruptions handled by the service, so you manage one connection instead of three. Kawl Now runs on OpenAI’s GPT-Live voice model.