Frontdesk.ai is an open-source, self-hostable AI receptionist built on the AssemblyAI Voice Agent API. It answers a business's calls, understands the caller, checks the real calendar, and books the appointment, then gives the owner a transcript, a recording, and a summary of every conversation. AssemblyAI is the core of the product. Each call is one Voice Agent session: speech recognition, turn detection, the language model, and the voice run in a single managed pipeline over one WebSocket, opened straight from the browser with a short-lived token. There is no separate speech-to-text, LLM, and text-to-speech stack to stitch together. Every session is tuned for that business. The business name goes in as a key term so proper nouns are transcribed correctly. The transcription prompt describes a reception call where callers spell out names and emails. Turn detection allows up to 2.5 seconds of silence, so callers are not cut off mid-spelling. Each tool has a 20 to 30 second timeout instead of the 120-second default. During the call the agent uses function tools that run inside the app: check_availability reads the calendar, book_appointment writes the booking, and search_knowledge answers from the business's own content. After hangup, a background worker pulls the session's timeline and audio from AssemblyAI and stores the transcript, recording, and summary. Around the receptionist is a complete front desk: services, hours, staff, calendar, clients, conversation history, analytics, and a public booking page. A business can scan its website to fill in services and knowledge automatically. Built with Next.js, Supabase (Postgres with Row Level Security), BullMQ, and Redis. Multi-tenant, runs with one Docker Compose command, and released under Apache 2.0. Customer conversations stay on infrastructure the business controls.
Category tags: