The tool reference.

make_call — place an outbound voice call

Pass to + a prompt and the assistant will speak it; the recipient's reply is recorded for up to 60 seconds by default. Or pass raw twiml for full control. Inline TwiML means no public webhook required.

Parameters

Example

// In your AI host's chat:
"use crixin voice to call +15555550100 and ask in Egyptian Arabic
what time their basketball court opens; record their reply for 60 seconds."

// What the agent calls:
make_call({
  to: "+15555550100",
  prompt: "السلام عليكم، هل ممكن تقولولي إيه مواعيد ملعب كرة السلة عندكم؟",
  voice: "Polly.Hala-Neural",
  language: "ar-EG",
  max_recording_seconds: 60
})

// Returns:
{
  "sid": "CA94e1d8…3a02",
  "status": "queued",
  "to": "+15555550100",
  "from": "+15555550100",
  "direction": "outbound-api",
  "start_time": null
}

get_call — look up one call by SID

Returns status, duration, price, start/end timestamps. Useful right after make_call to poll until the call moves to completed.

get_call({ call_sid: "CA94e1d8…3a02" })

// → status: 'completed', duration: '47', price: '-0.0085', price_unit: 'USD'

list_calls — recent call history

Filter by destination, source, status, or start window.

list_calls({
  status: "completed",
  start_time_after: "2026-05-01T00:00:00Z",
  page_size: 100
})

list_recordings — recordings attached to a call

Returns SIDs, durations, and direct media URLs. The media URLs require your Twilio auth header to fetch — they're not public.

list_recordings({ call_sid: "CA94e1d8…3a02" })

// →
[{
  sid: "RE5b8c…",
  call_sid: "CA94e1d8…3a02",
  duration: "47",
  status: "completed",
  media_url: "https://api.twilio.com/.../Recordings/RE5b8c….mp3"
}]

transcribe_call — Deepgram-powered transcription

Pulls the audio and runs it through Deepgram. 30+ languages including Arabic, Spanish, Portuguese, Mandarin. Returns transcript, confidence, word-level timestamps, and detected language. Requires DEEPGRAM_API_KEY in env.

transcribe_call({
  recording_sid: "RE5b8c…",
  language: "ar"
})

// →
{
  recording_sid: "RE5b8c…",
  language: "ar",
  duration_seconds: 47,
  transcript: "أهلاً، الملعب بيفتح الساعة سبعة الصبح وبيقفل الساعة عشرة بالليل…",
  confidence: 0.91,
  word_count: 38
}

send_sms — SMS / MMS

Optional messaging_service_sid for sticky-sender behavior across countries; optional media_url for MMS attachments.

send_sms({
  to: "+15555550100",
  body: "Reminder: 3pm appointment at Mansoura Stadium basketball court."
})

TTS voices — neural + standard

Twilio routes the voice parameter to Amazon Polly or Google TTS. Pick a neural voice when available — they're noticeably more natural.

Inline TwiML — full control without webhooks

Pass raw TwiML XML in the twiml parameter when the prompt/record flow isn't enough — multi-step menus, DTMF input, conferencing, etc.

make_call({
  to: "+15555550100",
  twiml: `<?xml version="1.0" encoding="UTF-8"?>
<Response>
  <Say voice="Polly.Joanna-Neural">Press 1 for hours, 2 for prices.</Say>
  <Gather numDigits="1" action="https://example.com/menu">
    <Say>Waiting...</Say>
  </Gather>
</Response>`
})

For multi-step flows you'll need a public callback (action="…"). For single-shot prompts, the built-in prompt + record covers most use cases.

Error handling — what goes wrong, and why

Every tool returns either a JSON payload or a structured MCP error with isError: true. The error message is the underlying Twilio (or Deepgram) error verbatim — easy to grep, easy to retry.

// Common shapes:
"Error: TwilioApiError: POST /Calls.json failed: { "code": 21211, "message": "Invalid 'To' Phone Number" }"
"Error: Deepgram transcribe failed (401): Invalid credentials"
"Error: Missing TWILIO_ACCOUNT_SID. Set it in your environment or Doppler."