ElevenLabs Agents

Interact with ElevenLabs Conversational AI agents to initiate voice conversations, retrieve call transcripts, and manage outbound calling capabilities.

Prerequisites

This skill requires the ElevenLabs Agents MCP server to be deployed and configured:

1. Deploy the MCP server from this repository's mcp-server/ directory to Vercel
2. Set the ELEVENLABS_API_KEY as a Vercel environment variable (not in .env files)
3. Add to your Claude Code MCP config (~/.claude.json or project .claude/settings.json):

{
  "mcpServers": {
    "elevenlabs-agents": {
      "url": "https://your-vercel-deployment.vercel.app/mcp",
      "transport": "http"
    }
  }
}

See the repository README for detailed setup instructions: https://github.com/taskcrew/elevenlabs-agents

When to Use This Skill

Use this skill when you need to:

• Initiate conversations with ElevenLabs AI agents
• Make outbound phone calls using AI agents
• Retrieve and analyze conversation transcripts
• Monitor call status and completion
• List available agents or phone numbers
• Build or test voice AI applications

Available Tools

Agent Management

• list_agents: Get all available ElevenLabs agents in your account
• get_agent: Retrieve detailed information about a specific agent

Conversation Initiation

• initiate_agent_call: Start a conversation with an agent and wait for completion
  • Default timeout: 50 seconds (configurable up to 300 seconds)
  • Returns full transcript upon completion
  • Best for quick, synchronous interactions

Outbound Calling

• list_phone_numbers: Get all phone numbers available for outbound calls
• initiate_outbound_call: Make a phone call to any number using an agent
  • Requires agent_id and to_number (E.164 format, e.g., +15551234567)
  • Returns conversation_id for tracking
  • Best for asynchronous phone calls

Status & Transcripts

• get_call_status: Check the current status of a conversation/call
  • Statuses: initiated, in-progress, processing, done, failed
  • Returns call_outcome: completed, failed, no_answer, or in_progress
• get_call_transcript: Retrieve full transcript and metadata from any finalized call
  • Works for completed calls (status: done/completed)
  • Works for failed calls (status: failed)
  • Works for no-answer calls (status: initiated with null accepted_time_unix_secs)
  • Returns conversation turns, timestamps, LLM usage, cost data, and call_outcome

Common Workflows

Quick Agent Testing

1. List agents to find the agent_id
2. Use initiate_agent_call with a test message
3. Review the returned transcript

Outbound Calling

1. List phone numbers to verify availability
2. Initiate outbound call with agent_id and phone number
3. Poll get_call_status to monitor progress
4. Retrieve transcript with get_call_transcript when done

Conversation Analysis

1. Get conversation status to check completion
2. Retrieve transcript with full metadata
3. Analyze conversation turns, timestamps, and metrics
4. Review LLM usage and cost data

Important Notes

Status Values

• initiated: Call set up but not started (may indicate no-answer if never accepted)
• in-progress: Active conversation
• processing: Call completed, being processed
• done: Fully completed and processed
• failed: Call encountered an error

Call Outcomes

The MCP adds a call_outcome field to help identify the result:

• completed: Call successfully completed (status: done/completed)
• failed: Call failed with an error (status: failed)
• no_answer: Call was initiated but never answered (status: initiated, accepted_time_unix_secs: null)
• in_progress: Call is still ongoing (status: in-progress/processing)

Phone Number Format

Always use E.164 format for phone numbers:

• Correct: +15551234567
• Incorrect: (555) 123-4567, 555-123-4567

Transcript Availability

Transcripts are available for:

• Completed calls: Full conversation with all turns
• Failed calls: Partial data with error information
• No-answer calls: Empty transcript with metadata showing the call was never answered
• In-progress calls: Will return a status message, check back later

Timeout Considerations

• Default timeout for initiate_agent_call: 50 seconds
• Maximum timeout: 300 seconds (Vercel serverless limit)
• For longer conversations, use the outbound calling pattern with status polling

Response Structure

Conversation Transcript

Each transcript contains:

• role: "agent" or "user"
• message: The spoken content
• time_in_call_secs: Timestamp within conversation
• conversation_turn_metrics: Performance metrics
• llm_usage: Token counts and costs
• source_medium: "audio" or "text"

Metadata

Full conversation metadata includes:

• Call duration and timestamps
• Cost breakdown (LLM charges, call charges)
• Phone call details (if outbound)
• Termination reason
• Features usage
• Analysis results (if enabled)

Example Usage

Test an agent with a simple message

Use initiate_agent_call with:
- agent_id: "agent_xxx"
- initial_message: "Hello, can you tell me about your services?"

Make an outbound sales call

1. List phone numbers to get phone_number_id
2. Use initiate_outbound_call with:
   - agent_id: "agent_xxx"
   - to_number: "+15551234567"
3. Poll status every 5-10 seconds
4. Get transcript when status is "done"

Retrieve a specific conversation

Use get_call_transcript with:
- conversation_id: "conv_xxx"

Related Documentation

For detailed API reference, search the indexed NIA documentation:

• "ElevenLabs conversation API"
• "ElevenLabs agents platform"
• "ElevenLabs outbound calling"