Embeddable voice agents

Give your customers an expert by their side

Marlin embeds in one script tag. Users talk, the agent reads your UI, highlights what to click, and responds out loud — no chat widget required.

Get started Log in See capabilities

New to Marlin? Create a free account — or tap Ava in the bottom-right to try it on this page.

1 script tag

to embed on any site

WebRTC

speech in and out

Page-aware

highlights & guidance

index.html

<!-- Add before </body> -->
<script
  src="https://cdn.onmarlin.com/widget.js"
  data-project="pk_live_your_project_key"
  data-api="https://onmarlin.com/api/session"
  data-target="#app"
  async
></script>

How it works

Understands your page, then acts on it

Marlin sends the agent a pruned accessibility tree with numbered element indices. The agent uses those indices to highlight, scroll, and guide — out loud.

Page context
Reads the live accessibility tree — roles, labels, values, and states — so the agent always knows what's on screen.
Element highlighting
Pulses a highlight on the exact element to click. Spoken guidance stays in sync with what users see.
On-demand snapshots
Captures a JPEG when the DOM isn't enough — charts, canvas content, and visual layouts.

page-context.jsonlive

[12]Submit order

buttondisabled

[13]Email

textboxrequired

[14]Shipping policy

link

[15]Voice sessions chart

canvasvisual

Realtime voice

Speech in and out over WebRTC. Users talk naturally and hear answers immediately.

Private by design

Ephemeral API keys, no screen-share prompt, in-memory state only.

Framework-agnostic

Scope the widget to any element. Works on static sites, SPAs, and multi-page apps.

analytics — voice sessionssample data

Visual understanding

Sees charts and canvas when the DOM falls short

Canvas and chart content isn't in the accessibility tree. Marlin captures a snapshot on demand, so the agent can describe trends and visual state — not just buttons and links.

Example response

“Voice sessions climbed through Q2 — June hit 91k. Want me to highlight the export button?”

Ask Ava about the chart on this page to see it in action.

Setup

Three steps to production

Copy the install snippet above, add a token endpoint on your server, and Marlin handles the rest.

Embed the script

Add the Marlin widget before </body>. Point data-api at your token endpoint.

Mint ephemeral keys

Expose POST /api/session on your server. Never ship your OpenAI key to the browser.

Ship voice guidance

Users talk, Marlin reads the page, highlights elements, and responds out loud.

app/api/session/route.tsView live endpoint →

// POST /api/session → ephemeral Realtime API key
export async function POST() {
  const session = await openai.realtime.sessions.create({
    model: "gpt-realtime-2",
  });
  return Response.json(session.client_secret);
}

Back to install snippet View token endpoint

Compare

Voice guidance, not another chat widget

Traditional support bots answer in a sidebar. Marlin reads your page, highlights elements, and guides users out loud.

Capability	Marlin	Intercom	Drift	Zendesk
Real-time voice (speech in / out)	Yes	Limited	No	No
Live page & DOM awareness	Yes	No	No	No
On-screen element highlighting	Yes	No	No	No
One-script embed install	Yes	Partial	Partial	Partial
No screen-share permission	Yes	No	No	No
Text chat	Out of scope	Yes	Yes	Yes