Client-facing voice agent demo
Click once, speak naturally, and compare six voice routes from cheapest demo mode to premium production voice. Every route can run on a website now and can be connected to phone later.
Select one of six routes, play English/German, start a conversation, and compare cost against quality.
Starter Voice
Allow microphone and speak. The agent will answer, then listen again.
Ready for hands-free web conversation. This is not real phone/SIP yet.
Sorted from cheapest internal cost to highest quality. Public prices are client-facing; real route costs stay behind your PIN.
This is the honest technical map. The code is ours, but the AI models are not local; they run in provider clouds.
| Layer | Current choice | Where it runs | Paid/free status | In-house? |
|---|---|---|---|---|
| Website page | Static HTML/CSS/JS | Cloudflare Pages edge | Cloudflare free/low cost | Yes, our code |
| Browser/demo route | Browser SpeechRecognition + browser speech + Groq | User's browser | Cheapest; demo-safe, not production-safe enough | Partly, browser engine |
| Paid transcription route | OpenAI Whisper fallback; future production STT can be OpenAI/Deepgram/ElevenLabs Scribe | OpenAI cloud | Paid per usage | No |
| LLM brain | Groq llama-3.3-70b-versatile | Groq cloud | API usage; currently cheap/free-tier friendly | No |
| OpenAI Starter voice | OpenAI TTS-1, alloy | OpenAI cloud | Paid per generated audio | No |
| Business voice | OpenAI TTS-1, nova | OpenAI cloud | Paid per generated audio | No |
| OpenAI HD voice | OpenAI TTS-1-HD, onyx | OpenAI cloud | Paid per generated audio; middle premium route | No |
| Professional voice | ElevenLabs Flash/Rachel; OpenAI onyx fallback | ElevenLabs/OpenAI cloud | Paid provider route; ElevenLabs account status must be watched | No |
| Backend API | Cloudflare Worker | Cloudflare edge | Free/low cost until real traffic | Yes, our code |
| Usage tracking now | Browser localStorage | This visitor's browser | Free, demo only | Yes, but not production durable |
| Phone calls | Future LiveKit SIP/Telnyx or Retell route | Provider cloud + telecom network | Not connected on this page yet | No, only orchestration would be ours |
After the client chooses a package, we install one script on their website and configure their knowledge base, escalation email, and allowed domains.
Starter Voice
For now this opens an email draft and stores the request in this browser for demo tracking. Backend email capture is the next build step.
Enter PIN to see internal costs, margin, and demo usage.
Wrong PIN.
Variable cost/min is the estimated provider cost for one real conversation minute: roughly one minute of customer speech plus one minute of AI answer. It is not a fixed invoice price.
Monthly examples below are voice usage only at 120 conversation minutes/day. They do not include your profit, setup work, support time, or taxes.
Currency note: providers usually bill in USD. The EUR values here are practical sales estimates, not accounting-grade exchange-rate calculations.
Supabase Pro is optional at MVP stage. For a real production dashboard with manuals/RAG/transcripts/usage, budget about USD 25/month for one shared Supabase project.
That project can serve all six routes and multiple clients until traffic, data isolation, or compliance says we should split it.
Active today: Cloudflare Pages/Worker, browser speech APIs, Groq LLM, OpenAI Whisper/TTS, OpenAI HD route, and Professional ElevenLabs route if the Worker secret remains valid.
Not connected here yet: Azure TTS, Deepgram, LiveKit, Telnyx, Supabase RAG, durable billing, manuals/PDF upload, and real phone numbers.
Local/self-hosted: none of the AI models. Only the product code and UI are ours right now.
No usage yet.