Build with Double.

Double is an embeddable AI video sales agent. One script tag drops a launcher on your site; clicking it opens a live, face-to-face video call with your org's agent: no forms, no calendar, no wait. This page covers the embed, the browser JavaScript API, authentication, the full REST API, webhooks, and how the conversation is driven.

Quickstart

Add one <script> tag to your site. It renders a launcher bubble in the corner; when a visitor clicks it, Double opens a live video call with your org's live agent. No agent id is required. The server falls back to your live agent automatically.

<!-- Double embed -->
<script src="https://your-double-host/w.js" data-key="pk_live_xxxxxxxxxxxx" async></script>

<script>
  window.DoubleQueue = window.DoubleQueue || [];
  window.Double = window.Double || function(){ (window.DoubleQueue=window.DoubleQueue||[]).push(arguments); };
</script>
<script src="https://your-double-host/w.js" data-key="pk_live_xxxxxxxxxxxx" async></script>

<div id="double-call" style="height:80vh"></div>
<script src="https://your-double-host/w.js"
        data-key="pk_live_xxxxxxxxxxxx"
        data-inline="#double-call" async></script>

Double('mount', { target: '#double-call' });

Pin a specific agent (optional)

By default the widget uses your live agent. To pin a particular one, add data-agent:

<script src="https://your-double-host/w.js"
        data-key="pk_live_xxxxxxxxxxxx"
        data-agent="agt_..." async></script>

JavaScript API

Once the loader is on the page it exposes a single global function, window.Double(cmd, arg), that talks to the widget iframe over postMessage.

Example: identify a logged-in user and pass context

// Tell the agent who's on the page…
Double('identify', {
  email: 'dana@acme.com',
  name:  'Dana Ortiz',
  plan:  'growth'
});

// …and what they're looking at right now.
Double('context', {
  page:      'pricing',
  cartValue: 1490,
  planViewed: 'scale'
});

// Open the call from your own button.
document.querySelector('#talk-to-sales')
  .addEventListener('click', function(){ Double('open'); });

Authentication

Double uses two keys. Pick the one that matches where your code runs.

Where to find your keys

In the Dashboard → Settings. In local mode you can also read them from GET /v1/bootstrap (below), which requires no auth from localhost and auto-resolves the default org.

Example: authenticated REST call

curl https://your-double-host/v1/agents \
  -H "Authorization: Bearer sk_live_your_secret_key"

REST API

All endpoints live under /v1, speak JSON, and return errors as { error: { message, detail? } } with a 4xx/5xx status. Management routes use the secret key; public routes use the publishable key.

Bootstrap

Returns the org's keys, the active brain mode, and available models. No auth required from localhost (local mode auto-uses the default org). This is what the widget and dashboard call at boot.

curl https://your-double-host/v1/bootstrap

Response

{
  "org": {
    "id": "org_...",
    "name": "Acme, Inc.",
    "publishable_key": "pk_live_xxxxxxxxxxxx",
    "settings": { }
  },
  "brain_mode": "hosted",
  "public_url": null,
  "models": { }
}

Agents

An agent binds a profile (face + voice), a script, a persona, objectives, and guardrails into one configurable rep.

List agents

curl https://your-double-host/v1/agents \
  -H "Authorization: Bearer sk_live_your_secret_key"

Create an agent

curl -X POST https://your-double-host/v1/agents \
  -H "Authorization: Bearer sk_live_your_secret_key" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "Maya (Inbound rep)",
    "status": "draft",
    "profile_id": "prof_...",
    "script_id": "scr_...",
    "persona_prompt": "Warm, concise solutions engineer.",
    "greeting": "Hey! Want a quick look at how this works?",
    "language": "en",
    "objectives": [
      { "key": "email", "label": "Work email", "description": "Capture a work email", "required": true }
    ],
    "guardrails": [
      { "kind": "never", "rule": "Never quote discounts above 20%." }
    ],
    "widget_config": {
      "accent": "#2447F5",
      "launcherLabel": "Talk to Maya",
      "position": "bottom-right",
      "bookingUrl": "https://cal.com/acme/demo",
      "transferEnabled": true,
      "badge": true
    }
  }'

Agent JSON shape

Sync to the video engine

After editing persona / script / knowledge, sync to build or update the avatar persona(s) used at call time.

curl -X POST https://your-double-host/v1/agents/agt_.../sync \
  -H "Authorization: Bearer sk_live_your_secret_key"

Profiles

A profile is the face and voice of a rep: either a ready-to-use stock replica or a custom clone.

List stock replicas

curl https://your-double-host/v1/stock-replicas \
  -H "Authorization: Bearer sk_live_your_secret_key"

Returns { data: [{ replica_id, replica_name, thumbnail_video_url, thumbnail_image_url, model_name }] }.

Create a stock profile (ready immediately)

curl -X POST https://your-double-host/v1/profiles \
  -H "Authorization: Bearer sk_live_your_secret_key" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "Maya",
    "kind": "stock",
    "replica_id": "r79e1c033f",
    "thumbnail_url": "https://.../maya.png"
  }'

Knowledge

Ground the agent in your own material. Add sources by URL, raw text, or file upload; they're chunked and indexed for retrieval at call time.

Add a URL or text source

curl -X POST https://your-double-host/v1/knowledge/sources \
  -H "Authorization: Bearer sk_live_your_secret_key" \
  -H "Content-Type: application/json" \
  -d '{
    "type": "url",
    "name": "Pricing page",
    "url": "https://acme.com/pricing",
    "agent_id": "agt_..."
  }'

JSON body is { type: 'url'|'text', name, url|text, agent_id? }. To upload a file (pdf/md/txt/html) instead, send multipart/form-data with a file field plus optional name / agent_id. Omitting agent_id makes the source available to every agent in the org.

Upload a file

curl -X POST https://your-double-host/v1/knowledge/sources \
  -H "Authorization: Bearer sk_live_your_secret_key" \
  -F "file=@./one-pager.pdf" \
  -F "name=Product one-pager" \
  -F "agent_id=agt_..."

Search the index

curl "https://your-double-host/v1/knowledge/search?q=refund+policy&agent_id=agt_..." \
  -H "Authorization: Bearer sk_live_your_secret_key"

Returns { results: [{ chunk_id, source_id, source_name, heading, text, score }] }.

Sessions

A session is one live call. The widget creates it with your publishable key; management routes read it with the secret key.

Start a session

Authorized by the publishable key. agent_id is optional. Omit it to use your live agent.

curl -X POST https://your-double-host/v1/sessions \
  -H "Content-Type: application/json" \
  -d '{
    "pk": "pk_live_xxxxxxxxxxxx",
    "agent_id": "agt_...",
    "visitor": {
      "anon_id": "v_8f21",
      "identify": { "email": "dana@acme.com", "name": "Dana Ortiz" },
      "page": "pricing",
      "context": { "planViewed": "scale" }
    }
  }'

Response

{
  "session_id": "ses_...",
  "conversation_url": "https://rtc.double.app/...",
  "conversation_id": "c_...",
  "ws_url": "wss://your-double-host/ws?session_id=ses_...",
  "brain_mode": "hosted",
  "agent": {
    "name": "Maya",
    "greeting": "Hey! Want a quick look?",
    "widget_config": { },
    "profile_thumbnail": "https://.../maya.png",
    "demo_available": true
  },
  "disclosure": true
}

End a session

curl -X POST https://your-double-host/v1/sessions/ses_.../end \
  -H "Authorization: Bearer sk_live_your_secret_key"

GET /v1/sessions?limit=&agent_id= lists sessions; GET /v1/sessions/:id returns full detail including turns, events, and post-call analysis. Ending a call triggers the post-call analysis pipeline.

Analytics

curl "https://your-double-host/v1/analytics/overview?days=30" \
  -H "Authorization: Bearer sk_live_your_secret_key"

Returns { totals: { sessions, minutes, leads, meetings, demos_started }, rates: { completion, demo_start, outcome }, by_day: [...], top_objections: [...], recent_sessions: [...] }.

Agent intel

A frontier model works in the background and seeds the live (fast) model: a durable playbook per agent (grounded facts, FAQ, objection angles, do-not-claim rules — distilled from knowledge, the script, and recent call analyses), a pre-call brief per session, and a coach note between turns. The playbook rebuilds itself when knowledge or the persona changes; rebuild on demand to review it.

curl -X POST https://your-double-host/v1/agents/agt_.../intel/rebuild \
  -H "Authorization: Bearer sk_live_your_secret_key"

Returns { playbook: { facts, differentiators, faq, objection_angles, do_not_claim }, built_at, stale }. Session briefs and coach notes appear in each session's event stream as intel.brief / intel.coach.

Billing & trial

Plans: free, starter, growth, scale. The 14-day trial requires a card, includes 30 minutes of live video time, and converts automatically when it ends. Exceeding a limit returns HTTP 402.

GET /v1/billing returns { plan, status, trial: { ends_at, days_left, video_minutes }, video_budget: { limit, used, remaining }, usage, limits, plans }. Card details on the trial endpoint are validated server-side; only brand, last4, and expiry are ever stored. Wire a payment processor (Stripe Elements) before charging real cards.

Workspace & team

Webhooks

Register webhook endpoints per org via POST /v1/integrations to receive server-side events. Double sends an HTTP POST with a JSON body shaped { event, ts, data }.

{
  "event": "session.ended",
  "ts": "2026-07-04T18:03:22.184Z",
  "data": {
    "session_id": "ses_...",
    "agent_id": "agt_..."
  }
}

Events

MCP server

Everything in the product is also exposed to AI agents over the Model Context Protocol — one endpoint, JSON-RPC over HTTP POST, with 45+ tools spanning agents, faces, scripts, knowledge, demos, sessions, leads, analytics, billing, and workspace admin. Every tool executes through the same REST API documented above, so behavior, validation, and limits are identical.

Connecting (OAuth — recommended)

OAuth-capable MCP clients need only the URL; discovery, dynamic client registration, and the Authorization Code + PKCE (S256) flow are automatic. You sign in and pick a workspace on the consent screen.

{
  "mcpServers": {
    "double": {
      "url": "https://your-double-host/mcp"
    }
  }
}

Connecting (manual bearer token)

For scripts and advanced setups, pass your org's secret key directly — it carries every scope.

{
  "mcpServers": {
    "double": {
      "url": "https://your-double-host/mcp",
      "headers": {
        "Authorization": "Bearer sk_live_your_secret_key"
      }
    }
  }
}

Scopes & tokens

OAuth tokens are scoped per domain — agents:read, agents:write, knowledge:write, sessions:read, billing:write, simulate:run, and so on. tools/list only shows tools your token can call; read-only tools never require a write scope. Access tokens live 24 hours; refresh tokens rotate on every use and reusing a rotated token revokes the whole family. Mutating tool calls land in the workspace audit log.

Discovery endpoints

Brain modes

The value of bootstrap.brain_mode tells you which engine drives the conversation. It is auto-detected. You don't set it directly.

hosted: default, zero-config

The avatar vendor's model drives the conversation, grounded in the persona, script, and knowledge that Double bakes into the session at start. Nothing else is required. This is the mode you get out of the box.

byo-llm: our policy loop drives every turn

When the server has a public HTTPS URL, the avatar pipeline calls Double's OpenAI-compatible bridge at /v1/llm/chat/completions on every turn. That routes the conversation through our policy loop (the script state machine, retrieval over your knowledge, guardrails, and tools), so Double drives the call rather than the vendor model alone.