Voice MCP

Let an AI agent place real phone calls and talk for you. Two interchangeable voice pipelines behind two MCP servers, plus a browser test page so you can try both without a phone line.

What it does

You give the agent a phone number and a goal in plain language ("Confirm tomorrow's 14:30 appointment, or move it to any morning slot"). It dials, holds the conversation in the callee's language (Slovenian works well), hangs up when the goal is met, and returns a transcript plus a short summary and an outcome.

Both servers speak the standard Model Context Protocol over streamable HTTP, so any MCP client (Claude, Cursor, Hermes, your own code) can use them.

Connect an MCP client

The MCP endpoints are private on purpose: a tool that can dial phone numbers costs money and must not be open to the internet. They listen on the server's loopback interface only and you reach them through an SSH tunnel (or run your client on the same host). Ask the owner for the tunnel command; it forwards a local port to each server.

Once the tunnel is up, point your MCP client at the local URL, path /mcp, transport streamable HTTP:

{
  "mcpServers": {
    "voice":   { "url": "http://localhost:<voice-port>/mcp" },
    "livekit": { "url": "http://localhost:<livekit-port>/mcp" }
  }
}

Or test from a terminal with the MCP inspector: npx @modelcontextprotocol/inspector --transport http --server-url http://localhost:<port>/mcp

voice-mcp tools (Grok Voice)

voice_call_dial parameters

Example

voice_call_dial({
  "to": "+38640123456",
  "goal": "Ask if the table for 4 on Friday at 19:00 is still booked. If not, book it.",
  "context": "Booking is under the name Novak.",
  "language": "sl"
})

→ {
  "ok": true,
  "outcome": "success",          // success | partial | failed | no_answer | timeout | error
  "summary": "Table confirmed for Friday 19:00, 4 people.",
  "transcript": [ {"role": "assistant", "text": "..."}, {"role": "user", "text": "..."} ],
  "duration_s": 41.2,
  "hung_up_by": "remote"
}

livekit-mcp tools (Soniox + LLM)

livekit_call_dial parameters

Example

livekit_call_dial({
  "to": "+38640123456",
  "goal": "Order one large margherita for pickup at 19:30 under the name Jakob.",
  "language": "sl",
  "open_with_greeting": true,
  "vocabulary": ["margherita"]
})

The result has the same shape as voice_call_dial: outcome, summary, transcript, duration_s, hung_up_by.

Call flow

1. The client calls *_call_dial with a number and a goal. The server checks that no other call is running.
2. The server dials through the SIP trunk. On pick-up, audio is bridged to the AI.
3. The agent talks until the goal is met, the callee hangs up, it decides to end the call, or max_seconds expires.
4. A summary and outcome are generated and the tool returns them, together with the full transcript. The call is saved to history.

Cost per minute

Rough list prices in USD at the time of writing; providers change them. Telephony is billed per second on the Standard plan.

Try it in the browser

No phone number and no MCP client needed: talk to either agent through your microphone. The test page is password protected, runs at most 10 minutes per session, one session at a time, and a small hourly limit.

Privacy: test sessions are not recorded. Phone calls made through the MCP servers keep a transcript in call history on the host.