Web app + API (Vercel)
A React + Vite single-page app, with serverless API functions for everything the dashboard and onboarding need.
Architecture overview
This section is for technical readers. It explains the shape of CoveCall — the services, what each is responsible for, and how a call flows — at a level that’s safe to publish.
The pieces
Section titled “The pieces”Voice engine (Render)
A real-time Node service that handles live call audio and runs the conversation — transcription, the language model, and speech.
Database & auth (Supabase)
Postgres for all data, with authentication and row-level security so each business sees only its own records.
Telephony (Twilio)
Phone numbers and live, two-way call audio via Media Streams; SMS for alerts and follow-ups.
How a call flows
Section titled “How a call flows”Caller
TwilioMedia Streams
Voice engineASR → LLM → TTS
Toolscalendar · CRM · SMS
Supabasecalls · summaries
The dashboard reads the results; the voice engine handles the live conversation.
- A caller dials your Twilio number; the audio is streamed to the voice engine.
- The engine runs the conversation — see the voice pipeline — following your agent configuration and Call Map.
- When the agent acts (book, look up a customer, send a text), it calls allow-listed tools that reach your connected integrations.
- Results — the transcript, summary, outcome, and any alerts — are written to Supabase, and the dashboard reads them.
Trust boundaries
Section titled “Trust boundaries”- The browser only ever talks to the API and reads data it’s authorized to see.
- The voice engine is reached through short-lived, signed authorization for each call — never a shared password in the browser.
- Connector credentials are encrypted and only readable by trusted server-side roles.
See Security & data for the details.
The voice pipelineASR → LLM → TTS and the latency budget.
Security & dataIsolation, encryption, privacy, and consent.
Reliability & statusHow the system stays up.