Matcha: AI Booking Platform for Appointments & Classes

Small service businesses like fitness studios, salons and clinics juggle bookings across messaging apps, calendars and payment links. As bookings grow, messages get missed, slots get double booked, and someone spends the evening chasing confirmations.

Customers want to book the way they already talk: a WhatsApp message. No app, no waiting for the front desk. The assistant has to understand Greek and English, offer only slots that are free, and never let two people take the last place. The booking page and the team calendar must agree with it at all times.

Matcha gives a business two ways to take bookings and one place to see them. A customer can message the business on WhatsApp, for example "any reformer pilates thursday evening?". The assistant answers, finds a free time and asks for confirmation before it books. Or the customer opens the business's own booking page and picks a slot. Both paths land on the same calendar, and confirmations and reminders go out automatically.

The assistant is a LangGraph agent. Model, tools and providers sit behind ports so each can change on its own. Its tools cover sessions, bookings, staff, locations, memberships and credit packs. Routine requests are matched by rules first, so about 40% of turns never call a model. Answers about hours, prices and availability are checked against the database before they go out. When a request is beyond it, the assistant hands over to a person.

The multi-tenant NestJS backend books with atomic MongoDB operations, so two requests cannot take the same place. When a place opens, the waitlist promotes the next person. Stripe handles payments and credit packs, and Google Calendar stays in sync. The dashboard updates live over SSE. Each tenant's data is kept apart, and every action is logged.

• - Live in early-access beta at my.bookmatcha.app: six months free, no card, Greek and English out of the box.
• - 50 parallel requests for a 10-place session: exactly 10 succeed.
• - Each channel has a latency budget. Timeouts, retries and circuit breakers keep every reply inside it.
• - Every agent change must pass a ~15,900-case test suite and a ~270-scenario eval harness.
• - Confirmations and reminders run on 38 scheduled jobs and an outbox, with Redis caching and MongoDB indexes behind them.