elevenlabs-agents-mcp-server

MCP server that lets Claude place outbound telephone calls via ElevenLabs Agents (Conversational AI) over a SIP trunk (e.g. Vonage) and report the outcome — status, success evaluation, summary, extracted data, and full transcript.

Tools

Typical flow: list_agents → list_phone_numbers → make_outbound_call → wait_for_call_completion → (optionally) get_call_transcript.

Related MCP server: Voice Call MCP Server

Setup

npm install
npm run build

Claude Desktop / Claude Code config

{
  "mcpServers": {
    "elevenlabs-agents": {
      "command": "node",
      "args": ["/absolute/path/to/elevenlabs-agents-mcp-server/dist/index.js"],
      "env": {
        "ELEVENLABS_API_KEY": "sk_..."
      }
    }
  }
}

Claude Code: claude mcp add elevenlabs-agents -e ELEVENLABS_API_KEY=sk_... -- node /path/to/dist/index.js

Vonage SIP trunk prerequisites (one-time, in dashboards — not via this server)

1. Vonage: create a SIP trunk / programmable SIP domain. For outbound-from-ElevenLabs you need a termination endpoint (hostname) that accepts SIP INVITEs from ElevenLabs, with digest auth credentials. Allow signaling from sip.rtc.elevenlabs.io (or your region's static SIP endpoint).

2. ElevenLabs → Phone Numbers → Import from SIP trunk: enter your Vonage number (E.164), outbound address (Vonage termination hostname), transport (TLS recommended), and SIP credentials.

3. Assign or note the agent; the imported number's phone_number_id is what elevenlabs_make_outbound_call uses as agent_phone_number_id.

4. Analysis fields (call_successful, summary, data collection) require evaluation criteria / data collection configured on the agent.

Environment variables

• ELEVENLABS_API_KEY (required)

• ELEVENLABS_API_BASE_URL (optional, default https://api.elevenlabs.io; use residency endpoints if applicable)

Notes

• make_outbound_call is asynchronous — the API returns as soon as the call is initiated. Post-call analysis is only available once status is done (there is a short processing window after hang-up).

• wait_for_call_completion caps at 900 s per invocation and tells you if it timed out so you can call it again.

• Dynamic variables passed to make_outbound_call are available in the agent's prompt as {{variable_name}}.