Bridge Format

​

Overview

1. show sponsored recommendations under assistant answers
2. let users start an in-session conversation with a brand agent

• frontend SDK for recommendation rendering and delegation lifecycle
• host-managed consent UI
• host-managed delegated runtime after activation

​

Integration Model

• recommendation selection
• sponsored recommendation payloads
• click and exposure tracking
• delegation eligibility
• delegation activation
• operator session lifecycle endpoints

• chat UI
• consent UX
• host-local delegated session state
• its own backend runtime
• model orchestration after delegation starts

​

Core Flow

​

1. Wrap chat with AdMeshProvider

import { AdMeshProvider } from 'admesh-ui-sdk';

<AdMeshProvider
  apiKey={process.env.NEXT_PUBLIC_ADMESH_API_KEY!}
  sessionId={sessionId}
  language="en-US"
  geo_country="US"
  userId={userId}
  model={chatModel}
  messages={messages}
>
  <YourChat />
</AdMeshProvider>

​

2. Render AdMeshRecommendations under assistant turns

import { AdMeshRecommendations } from 'admesh-ui-sdk';

{messages.map((msg) => (
  msg.role === 'assistant' && msg.userQuery && msg.userMessageId ? (
    <AdMeshRecommendations
      key={msg.messageId}
      messageId={msg.userMessageId}
      query={msg.userQuery}
      onDelegationConsent={handleDelegationConsent}
      onDelegationActivated={handleDelegationActivated}
    />
  ) : null
))}

​

3. User clicks Talk to Agent

• requests consent from the host
• calls AdMesh delegation activation
• returns a normalized delegation payload to the host

​

4. Host activates its delegated runtime

• delegation_session_id
• brand_name
• brand_id
• mcp_endpoint
• required_consent_types
• allowed_data_types
• consent_requirements
• supported_actions

​

5. User messages are routed through the brand agent

• the host backend routes the next turns through the brand runtime
• the host model can attach the brand MCP endpoint as a tool
• the host continues to own the final user experience

​

6. Stop the session

• the host calls stopDelegationSession(...) from admesh-ui-sdk
• the host clears its own local delegated state

​

Consent Handling

• SDK provides the delegation payload
• host renders its own consent dialog
• host decides how long consent is remembered

• once per brand per chat session

• session_goal
• delegate_reason
• eligibility_reason
• allowed_data_types
• required_consent_types
• consent_requirements

​

Mid-Conversation Data Collection

1. user asks for a high-intent action
2. LLM calls the brand tool/runtime
3. brand responds with a structured data request
4. host renders a form inside the conversation
5. user fills fields and confirms consent
6. host submits structured data back to the delegated runtime
7. brand flow continues

• do not rely on free-text collection for sensitive fields when a form is possible
• let the host UI collect email, phone, address, and consent explicitly

​

Recommended Host Callbacks

const handleDelegationConsent = async (recommendation: any) => {
  // Render your own consent dialog here.
  return true;
};

const handleDelegationActivated = async (payload: any) => {
  // Store host-local delegated state here.
  // Register MCP or brand runtime here.
};

const { stopDelegationSession } = useAdMesh();

await stopDelegationSession('user_stopped', {
  source: 'your-platform',
});

​

Architecture Boundary

• the AI platform should use admesh-ui-sdk for AdMesh/operator calls
• the AI platform may call its own backend for host-local delegation state

• AdMesh API calls: through the SDK
• host-local runtime calls: through the platform’s own backend

​

UX Recommendations

• Do not show recommendations on turns that were generated during an active delegated brand session.
• Once the delegated session is stopped, resume recommendations only for new post-stop user queries.
• Show clear provenance when a response came from a delegated brand agent.
• Use a branded consent dialog instead of window.confirm.

​

Next Steps

• Use React to install the SDK
• Use Tail & Product Format to render recommendation units
• Use Platform Overview to choose your integration model