AI voice agent that joins Zoom meetings via PSTN dial-in. Listens, transcribes, and speaks with AI-generated responses.
You → "Join my Zoom" → Agent dials Zoom PSTN → DTMF joins meeting
→ Live transcription via Telnyx webhooks
→ AI responses via GPT-4o-mini → Telnyx TTS
# 1. Install dependencies
npm install
# 2. Copy and fill in your credentials
cp .env.example .env
# 3. Join a meeting
node m3-voice-agent.js -m 83914076399 -p 953856node m3-voice-agent.js [options]
-m, --meeting-id <id> Zoom meeting ID (required)
-p, --passcode <code> Meeting passcode
-d <seconds> Max duration (default: 600)
| Variable | Required | Description |
|---|---|---|
TELNYX_API_KEY |
✅ | Telnyx API key |
TELNYX_DID |
✅ | Your Telnyx phone number |
TELNYX_CONNECTION_ID |
✅ | Call control application ID |
OPENAI_API_KEY |
✅ | OpenAI API key (for GPT responses) |
ZOOM_DIAL_IN |
Zoom dial-in number (default: +16699009128) | |
AGENT_NAME |
Display name (default: "AI Assistant") | |
AGENT_ROLE |
Role description | |
AGENT_INSTRUCTIONS |
Custom system prompt | |
BUFFER_DELAY |
Ms to wait before responding (default: 1500) | |
NO_SPEAK |
Set to "true" for listen-only mode | |
TRANSCRIPT_FILE |
Path to save transcript after call | |
USE_OPENCLAW_BRAIN |
Set to "true" to route responses via OpenClaw gateway | |
OPENCLAW_GATEWAY |
OpenClaw gateway URL (default: http://localhost:18789) | |
OPENCLAW_TOKEN |
OpenClaw API token (if auth enabled) |
| File | Purpose |
|---|---|
m3-voice-agent.js |
Main agent — full voice loop (transcribe + respond + speak) |
m2-live-transcribe.js |
Transcription-only mode (no AI responses) |
gemini-live-agent.js |
Gemini Live API integration (experimental — audio bridge WIP) |
bridge.js |
WebSocket bridge (requires public URL for media streaming) |
server.js |
Legacy Retell AI integration |
- Dial — Telnyx PSTN call to Zoom dial-in number
- Join — DTMF sequence: meeting ID → skip participant ID → passcode
- Tunnel — ngrok exposes local webhook server for Telnyx events
- Transcribe — Telnyx real-time transcription (Engine B)
- Think — OpenClaw brain (if enabled) or GPT-4o-mini generates response
- Speak — Telnyx TTS speaks response into the call
- ~15s for Zoom IVR greeting
- ~13s for DTMF sequence (meeting ID + passcode)
- ~1.5s buffer before responding (configurable)
- Total join time: ~30s
- Node.js 18+
- ngrok installed and authenticated
- Telnyx account with:
- A phone number (DID)
- Call control application (outbound channel limit ≥ 2)
- OpenAI API key
- Create a Call Control Application
- Set outbound channel limit to at least 2
- Assign your phone number to the application
- Note the connection ID — that's your
TELNYX_CONNECTION_ID
The webhook URL is set dynamically at runtime via the API (no manual config needed).
The agent responds in the same language the speaker uses:
- English → English response
- Chinese (Mandarin) → Chinese response
- Mixed → dominant language
Telnyx TTS supports en-US and cmn-CN voices.
- Voice quality: Telnyx basic TTS (robotic). Upgrade path: OpenAI TTS → audio streaming.
- Latency: ~2-4s round-trip (transcription + GPT + TTS). Reducible with streaming.
- One call per instance: Run multiple instances for concurrent meetings.
- PSTN only: No Zoom SDK integration (yet). Phone audio quality.
MIT