automateyournetwork/workspace/skills/twilio-inbound-voice

Twilio Inbound Voice Skill

Feature: 042-twilio-voice-mcp Priority: P2 User Stories: US5 (Inbound Conversations), US6 (Voice Commands)

Purpose

Enable John to call NetClaw's Twilio number and have a bidirectional voice conversation for status queries and network commands.

Trigger Conditions

This skill activates when:

1. Inbound Call Received
   • Call from whitelisted number ([WHITELIST_NUMBER])
   • Caller ID authenticated

Architecture

John's Phone
    │
    ▼
Twilio (Inbound Call)
    │
    ▼
Webhook: POST /webhooks/twilio/voice
    │
    ├─► Authenticate Caller ID
    │       └─► Not whitelisted? → Reject with message
    │
    ├─► Generate Greeting TwiML
    │       └─► "Hello John, this is NetClaw. How can I help?"
    │
    └─► Start Media Stream
            │
            ▼
    WebSocket: /webhooks/twilio/media-stream
            │
            ├─► Audio In → Whisper STT → Text
            │
            ├─► Text → Claude → Response
            │
            └─► Response → Twilio TTS → Audio Out

Workflow

Inbound Call Flow

Call Arrives at Twilio Number
    │
    ├─► Twilio sends webhook to NetClaw
    │
    ├─► Authenticate caller against whitelist
    │       └─► Rejected? → Play rejection message, hangup
    │
    ├─► Play greeting via TwiML <Say>
    │
    ├─► Connect Media Stream for real-time audio
    │
    ├─► [Loop] Conversation
    │       │
    │       ├─► Receive audio chunk
    │       │
    │       ├─► Transcribe via Whisper API
    │       │
    │       ├─► Process query/command
    │       │       └─► Network queries, status checks, commands
    │       │
    │       ├─► Generate response
    │       │
    │       └─► Speak via Twilio TTS
    │
    └─► On hangup: Log conversation to Memory MCP

Webhook Endpoints

POST /webhooks/twilio/voice

Initial call handler - returns TwiML.

Request: Twilio form-encoded data (CallSid, From, To, etc.)

Response (authenticated):

<?xml version="1.0" encoding="UTF-8"?>
<Response>
  <Say voice="Polly.Matthew">Hello John, this is NetClaw. How can I help you?</Say>
  <Connect>
    <Stream url="wss://your-server/webhooks/twilio/media-stream" />
  </Connect>
</Response>

Response (rejected):

<?xml version="1.0" encoding="UTF-8"?>
<Response>
  <Say voice="Polly.Matthew">This number is not authorized to call NetClaw. Goodbye.</Say>
  <Hangup />
</Response>

WebSocket /webhooks/twilio/media-stream

Real-time bidirectional audio stream.

Inbound Events:

• start: Call connected, media format info
• media: Audio chunk (mulaw, 8kHz, base64)
• stop: Call ended

Outbound Events:

• media: Response audio chunk

Voice Commands (US6)

John can issue network commands during the call:

Query Commands (Safe)

• "What's the status of my network?"
• "Are there any active incidents?"
• "Show me BGP neighbors on router-1"
• "What's the CPU on the firewall?"

Action Commands (Require Confirmation)

• "Run a health check on the core routers"
• "Restart the monitoring agent"
• "Clear the interface counters"

Confirmation Flow

John: "Restart the monitoring agent"
NetClaw: "You want me to restart the monitoring agent. Say 'yes' to confirm or 'no' to cancel."
John: "Yes"
NetClaw: "Restarting monitoring agent... Done. Agent is now running."

Guardrails

• Caller ID Authentication: Must be in whitelist
• Content Sanitization: Response sanitized before TTS
• Command Confirmation: Impactful actions require verbal "yes"
• Session Timeout: Max 15 minutes per call
• Logging: Full transcript logged to Memory MCP

Message Templates

Greeting

Hello John, this is NetClaw. How can I help you?

Status Response

Your network status:
{device_count} devices monitored.
{healthy_count} operational.
{incident_count} active incidents.
Would you like more details?

Confirmation Request

You want me to {action_description}.
Say 'yes' to confirm or 'no' to cancel.

Error Response

I'm sorry, I couldn't complete that request.
Error: {error_summary}.
Would you like to try something else?

Goodbye

Goodbye John. This conversation has been logged.

Integration Points

• Twilio Media Streams: Real-time audio
• OpenAI Whisper API: Speech-to-text
• Twilio TTS: Text-to-speech (Polly voices)
• Memory MCP: Conversation logging
• All MCP servers: Query execution

Security

1. Signature Validation: All webhooks validated via Twilio signature
2. Whitelist Authentication: Caller ID must be whitelisted
3. Content Sanitization: No IPs, credentials in spoken output
4. Audit Trail: Full transcript logged

Example Conversation

[Call connects]
NetClaw: "Hello John, this is NetClaw. How can I help you?"

John: "What's the status of my network?"

NetClaw: "Your network status: 47 devices monitored. 45 operational,
         2 with warnings. No active incidents. The warnings are on
         switch-3 and ap-lobby-1 due to high memory usage."

John: "Run a health check on switch-3"

NetClaw: "You want me to run a health check on switch-3.
         Say 'yes' to confirm or 'no' to cancel."

John: "Yes"

NetClaw: "Running health check on switch-3...
         Check complete. CPU: 23%, Memory: 78%,
         All interfaces up. The high memory is within normal range."

John: "Thanks, goodbye"

NetClaw: "Goodbye John. This conversation has been logged."
[Call ends]

Success Criteria

• [ ] Inbound calls authenticated within 5 seconds
• [ ] Unauthorized callers rejected with message
• [ ] Voice queries answered accurately
• [ ] Commands require verbal confirmation
• [ ] Full transcript logged to Memory MCP
• [ ] Max 15-minute call duration enforced