Phone Call Skill

Make autonomous phone calls with a goal-driven AI agent. The AI handles the conversation until the goal is achieved.

Prerequisites

1. Required configuration:

   concierge config set twilioAccountSid <your-sid>
   concierge config set twilioAuthToken <your-token>
   concierge config set twilioPhoneNumber <your-number>
   concierge config set deepgramApiKey <your-key>
   concierge config set elevenLabsApiKey <your-key>
   concierge config set elevenLabsVoiceId <voice-id>
   concierge config set anthropicApiKey <your-key>
   

2. Optional for auto-managed ngrok:

   concierge config set ngrokAuthToken <your-ngrok-token>
   

Usage

Basic call

concierge call "+1-555-123-4567" \
  --goal "Book a hotel room for February 15" \
  --name "John Smith" \
  --email "john@example.com" \
  --customer-phone "+1-555-444-1212" \
  --context "2 nights, king bed preferred"

Interactive mode

concierge call "+1-555-123-4567" \
  --goal "Make a reservation" \
  --name "John Smith" \
  --email "john@example.com" \
  --customer-phone "+1-555-444-1212" \
  --interactive

In interactive mode, you type what the AI should say in real-time.

Infrastructure behavior

• By default,

  call

  auto-starts

  ngrok

  and

  server

  if server is unavailable.
• Use

  --no-auto-infra

  to disable this and run everything manually.
• Auto-managed processes are stopped automatically when the call ends.
• Log files are written to:

  • ~/.config/concierge/call-runs/<run-id>/server.log

  • ~/.config/concierge/call-runs/<run-id>/ngrok.log

Server management

# Check server status
concierge server status

# Start server
concierge server start --public-url <ngrok-url>

# Stop server
concierge server stop

Preflight checks

Before dialing, the system validates:

• Local runtime dependencies (

  ffmpeg

  binary + MP3 decode support, plus

  ngrok

  if auto-infra is used)
• Twilio credentials/account status/from-number availability
• Deepgram API key/auth reachability
• ElevenLabs character quota sufficiency (estimated call budget)

How It Works

1. CLI sends a call request with goal + customer identity details
2. The server places the call via Twilio
3. Audio streams bidirectionally via WebSocket
4. Deepgram transcribes human speech in real-time
5. Claude generates appropriate responses
6. ElevenLabs synthesizes speech for responses
7. Call continues until goal is achieved or human hangs up

Examples

Book a hotel reservation

concierge call "+1-800-HILTON" \
  --goal "Book a room for 2 nights" \
  --name "Sarah Johnson" \
  --email "sarah@example.com" \
  --customer-phone "+1-555-000-2222" \
  --context "Check-in: March 10, Guest: Sarah Johnson, King bed, non-smoking"

Make a restaurant reservation

concierge call "+1-555-DINER" \
  --goal "Reserve a table for dinner" \
  --name "Garcia" \
  --email "garcia@example.com" \
  --customer-phone "+1-555-000-3333" \
  --context "Party of 4, 7:30 PM, Saturday, name: Garcia"

Cancel an appointment

concierge call "+1-555-DOCTOR" \
  --goal "Cancel appointment" \
  --name "Mike Chen" \
  --email "mike@example.com" \
  --customer-phone "+1-555-000-4444" \
  --context "Patient: Mike Chen, Appointment on Tuesday at 2 PM"

Supported Voice IDs

Some popular ElevenLabs voices:

• EXAVITQu4vr4xnSDxMaL

  - Rachel (default, conversational female)

• pNInz6obpgDQGcFmaJgB

  - Adam (conversational male)

• 21m00Tcm4TlvDq8ikWAM

  - Rachel (narration)

• AZnzlk1XvdvUeBnXmlld

  - Domi (young female)

Set your preferred voice:

concierge config set elevenLabsVoiceId <voice-id>

Latency

Target voice-to-voice latency: < 500ms

• Deepgram STT: ~150ms
• Response generation: ~100-200ms
• ElevenLabs TTS: ~75ms
• Network: ~50ms

Troubleshooting

Server won't start

• Check all config keys are set:

  concierge config show

• If using manual mode, ensure ngrok is running and URL is correct
• Check port 3000 is available

Call not connecting

• Verify Twilio phone number is active
• Check Twilio account has sufficient balance
• Ensure ngrok URL is publicly accessible (manual mode)

TTS fails mid-call

• Check ElevenLabs quota/credits.
• New preflight usually catches this before dialing.
• If it still happens, reduce prompt/context length or top up ElevenLabs.

Audio quality issues

• ElevenLabs uses optimized phone call settings
• Deepgram uses the phone call model
• Audio is at 8kHz (telephone quality)