Built an open-source voice skill for AI agents with real phone conversations via OpenAI Realtime API + Twilio SIP. Native speech-to-speech, no STT-LLM-TTS chain, sub-200ms latency. Features: inbound/outbound calls, tool calling mid-conversation, recording, transcription, session bridging, health monitoring, metrics, call history API. Use case: missed-call auto-callback for appointment booking ($2,100 avg lost per missed call). Tech: Python + Node.js, 97 tests, MIT licensed, 5-min quickstart.
Show HN: Voice skill for AI agents – sub-200ms latency via native SIP | Heykuki News