Your instinct is spot-on. Do not build this on raw LiveKit code if you cannot code yourself. While LiveKit is the gold standard for flexibility, using Claude Code to duct-tape an agent worker together is fine for a weekend experiment, but a liability on a live business phone line. The moment a SIP session drops, a turn-detection threshold fails, or a runtime error occurs during business hours, relying on an AI chat loop to debug production logs will quickly become overwhelming. ElevenLabs Conversational AI is significantly easier and gives you a clean interface. However, its native phone-call and SIP handling is still relatively basic compared to dedicated telephony platforms, and troubleshooting custom tool calls can become restrictive as the business demands grow. The Recommended Approach Rather than forcing yourself into raw code with LiveKit or hitting the telephony limitations of pure ElevenLabs, the best path is to treat ElevenLabs as the voice engine (TTS) inside a telephony-first voice orchestrator: 1. Use ElevenLabs purely for voice synthesis where it shines, keeping conversational Czech sounding natural. 2. Keep the orchestration layer no-code/low-code, so you have a dashboard to manage prompts, call recordings, latency settings, and SIP configurations without writing Python or TypeScript Two Crucial Recommendations (Outside What We Discussed) 1. Look into Retell AI or Vapi instead of raw LiveKit:These platforms are purpose-built for real phone calls and SIP trunking. They handle the hard telecom plumbing (interruption handling, silence detection, latency optimization, call transfers) and allow you to simply plug in your ElevenLabs API key for speech. You get full control over the STT model (like Deepgram or Whisper for Czech) and LLM without maintaining server code. 2. Use n8n (or Make) as your tool-calling engine: Instead of coding custom API endpoints for the bot to check availability or book appointments, route your voice agent's custom tools through a visual webhook automation platform like n8n. If a booking fails or a date is misformatted, you can visually trace and fix the logic in seconds without touching code.

@Jaroslav Záveský Totally understand your concern regarding DPAs and EU regulations—that’s a critical priority when dealing with European businesses. To clear up the confusion on both points: 1. ElevenLabs CAN run on a phone number directly: ElevenLabs Conversational AI has native telephony support. You don’t need an external platform just to get a phone number working. You can plug in an existing SIP trunk (from Telnyx or a local Czech provider) directly into the ElevenLabs dashboard without writing code. Where ElevenLabs is limited compared to dedicated telecom platforms is advanced call routing (like warm transfers to an office desk phone or complex IVR trees), but for a standard inbound receptionist that answers questions and logs bookings, ElevenLabs’ built-in telephony is more than capable. 2. What "telephony platforms" mean: They are the carriers and aggregators that connect internet AI audio to real phone lines (PSTN)—think Twilio, Telnyx, or orchestrators like Retell/Vapi. 3. The DPA / EU Compliance Reality: Platforms like Retell and Vapi actually do provide standard DPAs with EU Standard Contractual Clauses (SCCs). However, European businesses often scrutinize where the audio packets and transcripts travel. Since you don't want to manage code, your best and cleanest path is actually using ElevenLabs directly: - Build the agent inside ElevenLabs Conversational AI. - Connect it to a compliant SIP trunk provider (like Telnyx with EU data anchoring, or a registered Czech VoIP carrier). - Sign a Data Processing Agreement with ElevenLabs and your SIP carrier. That keeps your stack 100% no-code, handles the legal paperwork cleanly, and avoids building a custom server. If you run into any snags configuring the SIP trunk, prompt tuning the Czech conversational flow, or hooking up the booking calendar, give me a shout. I handle these technical voice builds regularly and can easily jump in to help set it up for your client so it runs smoothly.