📖 The AI Tool Bible
Workflow · 5 steps

Voice-agent production stack (phone → transcribe → reason → speak → hand-off)

Full voice-agent chain with a real phone number, real reasoning, and clean escalation to a human.

Ships a phone-answering agent in 2 hours. Every hop matters for latency.

  1. 1
    Vapi

    Vapi provisions a Twilio phone number + WebRTC session. This handles turn-taking + barge-in.

  2. 2
    Deepgram

    Deepgram Nova-3 real-time transcription (100-200ms latency).

  3. 3
    Claude

    Claude Sonnet 5 for reasoning. Include tool-use for CRM lookups, appointment booking, etc.

  4. 4
    ElevenLabs

    ElevenLabs v3 text-to-speech. Streams audio back through Vapi (~300ms to first audio).

  5. 5
    Vapi

    Escalate to a human via Twilio transfer when the agent hits configured triggers ("speak to a manager", 3 failed attempts, etc).