Embed the visitor widget / SDK

Three levels, from zero code to full control.

Level 0: a link

Publish a channel and send the URL. Works on phones; no app.

Level 1: an iframe

<iframe src="https://your-mynah:8443/channel/sales?k=TOKEN"
        allow="microphone; autoplay" style="width:100%;height:100vh;border:0"></iframe>

Restrict the channel to your domain in 发布管理 so nobody else can embed it. The allow attribute is required or the browser blocks the microphone inside the frame.

Level 2: the SDK

The browser SDK is served from your own cored at /sdk/mynah.js (global MynahSDK, minified /sdk/mynah.min.js; the old /sdk/personalive.js name still works) and also lives in web/sdk/ as an ES module. A runnable example ships at /sdk-example.html. It renders the avatar into any element, keys out the green screen onto a transparent canvas, and exposes events and abilities so the digital human can act inside your page.

<div id="stage" style="width:480px;height:720px"></div>
<script src="https://your-mynah:8443/sdk/mynah.js"></script>
<script>
  const human = new MynahSDK.Mynah({
    endpoint: "https://your-mynah:8443",
    slug: "sales",        // published channel
    token: "TOKEN",       // only for token-mode channels
    mic: true,            // ask for the microphone
    chroma: true,         // transparent canvas instead of green
    bg: "",               // or a background image URL
  });
  human.mount(document.getElementById("stage"));
  human.on("subtitle", ({ text }) => console.log("avatar:", text));
  human.on("user",     ({ text }) => console.log("visitor:", text));
  human.on("state",    (s) => console.log(s.state));   // idle | listening | thinking | speaking
  await human.connect();
  human.say("您好，欢迎光临。");          // speak verbatim
  human.ask("今天有什么优惠？");          // through the LLM + knowledge base
</script>

API

Method	Does
mount(el)	render into el (transparent canvas when chroma is on)
connect() / disconnect()	open / close the WebRTC session
say(text)	speak verbatim
ask(text)	answer through the configured brain; may trigger abilities
interrupt()	stop speaking now
mic(on)	mute / unmute the microphone track without renegotiating
on(event, cb) / off(event, cb)	events: connected, disconnected, state, subtitle, user, tool, error
ability(def) / removeAbility(name)	register a function the model can call

Abilities: let the avatar drive your page

human.ability({
  name: "goto_page",
  description: "Open a page on the big screen when the visitor asks to see a product line.",
  parameters: { type: "object", properties: { page: { type: "string" } }, required: ["page"] },
  blocking: false,
  run: ({ page }) => router.push(page),
});

The model sees the description, decides when to call it, cored relays the call to your page over the data channel, your run executes, and if blocking is true its return value is fed back into the answer (the avatar says a short filler while it waits). This is how a digital guide switches slides or looks up an order without any server-side change.

Custom element

The same thing declaratively:

<script type="module" src="/sdk/mynah.js"></script>
<mynah-avatar endpoint="https://your-mynah:8443" slug="sales" mic>
  <mynah-ability name="goto_page" description="Open a page on the big screen"
                       parameters='{"type":"object","properties":{"page":{"type":"string"}}}'></mynah-ability>
</mynah-avatar>

Level 3: raw HTTP

For non-browser clients the visitor surface is five routes: POST /offer (or /channel/offer with slug and token), /human, /interrupt_talk, /is_speaking, /action. The README lists the bodies. CORS is enabled on all of them.

Agents

An LLM agent is just another client. The DeepSeek Harness plugin wraps these routes as tools so an agent can make the avatar speak, wave and report.