Grok Customer Support

Grok Voice ↔ Twilio bridge hangup gate: hang up only after the errand success or a clear dead-end is confirmed; never mid-ID / mid-reference / on hold. The voice model may request hangup; the operator (or bot) must approve before the Twilio leg ends.

Outbound voice stack: Node bridge that connects Twilio Media Streams (μ-law 8 kHz) to xAI Grok Voice realtime (wss://api.x.ai/v1/realtime?model=grok-voice-latest, audio/pcmu). Place calls with POST /call; expose PUBLIC_HOST over WSS (e.g. ngrok) so Twilio can reach /media-stream.

Coach mode pattern: poll a live transcript (them vs agent), accept steer text from the operator in chat, POST /steer so Grok updates mid-call without announcing that it is being coached. Operator does not need to hear raw call audio.

Enable Twilio dual-channel recording on every outbound bridge call so recordings can be pulled by call SID from Twilio Monitor → Logs → Recordings.

Prefer a custom Grok Voice clone from the xAI console for outbound calls that should sound like the user. Grok Voice Agent Builder free numbers are mainly inbound; Google Voice is a poor fit for this bridge. Keep Twilio and xAI secrets in .env only — never in chat.

Hold-music / IVR: raise VAD threshold and silence duration so hold music does not trigger endless agent speech; cancel in-flight responses while hold music is detected.

Non-marketplace services this bot uses: xAI Grok Voice (realtime WebSocket API) and a self-hosted Node Twilio Media Streams bridge. Twilio marketplace plugin covers Twilio APIs; the bridge itself is local code, not an MCP server.

Twilio Skills and MCP provide procedural knowledge for AI coding agents — which APIs to use, in what order, and what to avoid. Covers Messaging, Voice, Verify, SendGrid, and 30+ products.