ElevenLabs + MCP Platform integration

Connect Cipher's MCP verification tools to your ElevenLabs conversational AI agents. Use Claude Desktop or custom MCP clients to add real-time caller verification to your voice workflows.

Part 1: API Setup

Configure the MCP server and middleware for programmatic verification

Set up Claude Desktop with Cipher MCP

First, configure Claude Desktop to connect to Cipher's MCP server. Install and configure the server with your Cipher API key.

claude_desktop_config.json

{
  "mcpServers": {
    "cipher-verification": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-cipher"],
      "env": {
        "CIPHER_API_KEY": "sk_live_your_key_here",
        "MCP_ENDPOINT": "https://api.getcipher.ai/mcp/sse"
      }
    }
  }
}

Design your verification conversation flow

In Claude Desktop, create a verification workflow that processes caller information from ElevenLabs conversations. Claude will use Cipher's MCP tools to verify identities and guide the conversation.

Example Claude conversation

You: An ElevenLabs agent just received this from a caller.
Session ID: "elevenlabs-conv-xyz789"
Caller: "Hello, my name is Michael Chen and I&apos;m calling
to check on my insurance claim. My policy number is
INS-445566 and I live at 789 Maple Drive, Austin Texas."

Claude: I'll verify this caller using Cipher.
[Uses verify_caller tool]

Verification Results:
- Name: Michael Chen (Confirmed)
- Address: 789 Maple Drive, Austin TX (Confirmed)
- Policy Number: INS-445566 (Recorded)
- Trust Score: 68%
- Status: Moderate confidence, phone verification recommended

Suggested response for ElevenLabs agent:
"Thank you, Mr. Chen. For security purposes, can you
please confirm the phone number on file for your account?"

Integrate with ElevenLabs WebSocket API

Use ElevenLabs' WebSocket API to stream conversation data to a middleware service that communicates with Claude Desktop via MCP. This enables real-time verification during conversations.

WebSocket connection

wss://api.elevenlabs.io/v1/convai/conversation

Message types to process

conversation_initiation_metadata - Start verification
user_transcript - Extract and verify caller statements
agent_response - Inject verification questions

Example middleware integration

// Listen to ElevenLabs WebSocket
ws.on('message', async (data) => {
  const msg = JSON.parse(data);

  if (msg.type === 'user_transcript') {
    // Send to Claude via MCP
    const verification = await claudeMCP.verifyCall({
      session_id: conversationId,
      caller_message: msg.transcript
    });

    // Update ElevenLabs agent context
    if (verification.trust_score < 80) {
      ws.send({
        type: 'add_to_agent_context',
        text: verification.next_question
      });
    }
  }
});

WebSocket integration flow

API integration architecture

ElevenLabs Agent (Conversation) → WebSocket Stream

↓

Your Middleware Server

↓

Claude Desktop (via MCP)

↓

Cipher MCP Server (verify_caller tool)

↓

Trust Score + Next Question

↓

Your Middleware → ElevenLabs (inject into agent context)

Part 2: Agent Platform Setup

Configure your ElevenLabs agent for verification-aware conversations

Create or select an ElevenLabs agent

Navigate to your ElevenLabs dashboard and create a new conversational AI agent or select an existing one to enhance with caller verification.

Configure the agent's system prompt

Update your agent's first message and system instructions to include verification guidelines. This helps the agent naturally collect identity information during conversations.

Example system prompt with verification

You are a customer service agent for [Company Name].

VERIFICATION PROTOCOL:
Before providing any sensitive information or making changes to accounts, you must verify the caller&apos;s identity by collecting:

1. Full name (required)
2. Complete address including street, city, state, and zip code (required)
3. Phone number associated with the account (required)
4. Date of birth or last 4 digits of SSN (for high-security requests)

CONVERSATION GUIDELINES:
- Begin by greeting the caller warmly
- Ask "How can I help you today?" to understand their request
- Naturally collect verification information during the conversation
- Be polite and explain that verification is for their security
- Wait for system confirmation of identity verification before proceeding with sensitive requests
- If verification fails, politely ask for additional information or offer to transfer to a supervisor

TRUST LEVELS:
- Below 60%: Continue collecting information, do not proceed with requests
- 60-79%: Handle basic inquiries only, no account changes
- 80%+: Fully verified, proceed with all authorized requests

Configure agent knowledge base

Add verification policies and procedures to your agent's knowledge base. This ensures consistent handling of verification across all conversations.

Example knowledge base entry

# Caller Verification Standards

## Required Information for Identity Verification

### Standard Verification (Most Requests)
- Full legal name
- Complete mailing address
- Primary phone number

### Enhanced Verification (Sensitive Operations)
All standard verification items PLUS one of:
- Date of birth
- Last 4 digits of Social Security Number
- Account PIN (if applicable)
- Security question answer

## Verification Process

1. **Collection Phase**: Gather information naturally during conversation
2. **System Verification**: Trust score calculated automatically
3. **Confirmation**: Wait for system confirmation before proceeding

## Trust Score Thresholds

- **0-59%**: Not verified - continue collecting information
- **60-79%**: Partially verified - basic inquiries only
- **80-100%**: Fully verified - proceed with all requests

## Special Cases

- **Failed Verification**: Offer to transfer to human agent
- **Suspicious Activity**: Note concerns and transfer immediately
- **Technical Issues**: Fallback to traditional verification methods

Set up conversation variables

Configure conversation variables to track verification state and collected information throughout the interaction.

verification_status

Current verification status: "pending", "partial", or "verified"

trust_score

Numerical trust score (0-100) from Cipher verification

collected_info

Object containing name, address, phone, and other collected details

Configure voice and personality

Select an appropriate voice from ElevenLabs' library that matches your brand and use case. Configure speaking rate and other voice parameters for optimal verification conversations.

Recommended voice characteristics

Clear and professional tone
Moderate speaking pace for information collection
Warm and friendly to put callers at ease
Authoritative for security-related messaging

Test your agent

Use ElevenLabs' built-in testing tools to have test conversations with your agent. Verify that it properly collects information and responds appropriately to different verification scenarios.

Test multiple scenarios: successful verification, partial verification, and failed verification. Monitor the Cipher dashboard to see trust scores and collected information in real-time.

Advanced configuration

Optional enhancements for production deployments

Custom MCP client integration

For fully automated verification without Claude Desktop, build a custom MCP client that directly integrates with ElevenLabs' API. This provides lower latency and tighter integration.

Benefits: No human in the loop, direct API integration, custom business logic

Multi-language support

Configure multiple agents for different languages, each with appropriate verification prompts and knowledge bases. Cipher supports verification in multiple languages.

Fallback handling

Configure graceful fallbacks for when verification fails or technical issues occur. Set up transfer-to-human workflows for edge cases.

Production best practices: Implement comprehensive logging, set up monitoring alerts for failed verifications, use A/B testing to optimize your verification prompts, and regularly review conversation transcripts to improve the verification flow.

Next steps

MCP Overview →

Understand Cipher's MCP architecture and tools

Vapi + MCP →

Integrate Cipher MCP with Vapi assistants

Proxy Platform →

Explore Cipher's proxy platform for voice agents

Dashboard →

Monitor verification activity across all platforms

{ "mcpServers": { "cipher-verification": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-cipher"], "env": { "CIPHER_API_KEY": "sk_live_your_key_here", "MCP_ENDPOINT": "https://api.getcipher.ai/mcp/sse" } } } }

You: An ElevenLabs agent just received this from a caller. Session ID: "elevenlabs-conv-xyz789" Caller: "Hello, my name is Michael Chen and I'm calling to check on my insurance claim. My policy number is INS-445566 and I live at 789 Maple Drive, Austin Texas." Claude: I'll verify this caller using Cipher. [Uses verify_caller tool] Verification Results: - Name: Michael Chen (Confirmed) - Address: 789 Maple Drive, Austin TX (Confirmed) - Policy Number: INS-445566 (Recorded) - Trust Score: 68% - Status: Moderate confidence, phone verification recommended Suggested response for ElevenLabs agent: "Thank you, Mr. Chen. For security purposes, can you please confirm the phone number on file for your account?"

// Listen to ElevenLabs WebSocket ws.on('message', async (data) => { const msg = JSON.parse(data); if (msg.type === 'user_transcript') { // Send to Claude via MCP const verification = await claudeMCP.verifyCall({ session_id: conversationId, caller_message: msg.transcript }); // Update ElevenLabs agent context if (verification.trust_score < 80) { ws.send({ type: 'add_to_agent_context', text: verification.next_question }); } } });

You are a customer service agent for [Company Name]. VERIFICATION PROTOCOL: Before providing any sensitive information or making changes to accounts, you must verify the caller's identity by collecting: 1. Full name (required) 2. Complete address including street, city, state, and zip code (required) 3. Phone number associated with the account (required) 4. Date of birth or last 4 digits of SSN (for high-security requests) CONVERSATION GUIDELINES: - Begin by greeting the caller warmly - Ask "How can I help you today?" to understand their request - Naturally collect verification information during the conversation - Be polite and explain that verification is for their security - Wait for system confirmation of identity verification before proceeding with sensitive requests - If verification fails, politely ask for additional information or offer to transfer to a supervisor TRUST LEVELS: - Below 60%: Continue collecting information, do not proceed with requests - 60-79%: Handle basic inquiries only, no account changes - 80%+: Fully verified, proceed with all authorized requests

# Caller Verification Standards ## Required Information for Identity Verification ### Standard Verification (Most Requests) - Full legal name - Complete mailing address - Primary phone number ### Enhanced Verification (Sensitive Operations) All standard verification items PLUS one of: - Date of birth - Last 4 digits of Social Security Number - Account PIN (if applicable) - Security question answer ## Verification Process 1. **Collection Phase**: Gather information naturally during conversation 2. **System Verification**: Trust score calculated automatically 3. **Confirmation**: Wait for system confirmation before proceeding ## Trust Score Thresholds - **0-59%**: Not verified - continue collecting information - **60-79%**: Partially verified - basic inquiries only - **80-100%**: Fully verified - proceed with all requests ## Special Cases - **Failed Verification**: Offer to transfer to human agent - **Suspicious Activity**: Note concerns and transfer immediately - **Technical Issues**: Fallback to traditional verification methods