Blackout curtains, pet fees, pillow menus, disability-friendly bed height — these are the details that decide a booking and that price-and-location search cannot express. Exposing `get_filters` as its own tool is the right call, because it lets the model discover valid values rather than inventing them. The payment split is worth noting for anyone building on top: the agent produces a link, the human pays, the booking confirms — the model never touches card data. There is also an affiliate programme paying 4%, 7% or 10% of booking value by monthly volume.
The official WinWin.travel MCP server: a hosted endpoint giving an agent access to hotel inventory the README puts at 3 Million+ properties, with search, filter discovery, reservation and status tools. Booking completes through a payment link the guest opens themselves, so card details never pass through the model.
- `search` — available room offers by destination, stay dates and guest composition, with optional filters covering meal plans, cancellation policies, amenities, accessibility features, pet policies and 500+ other criteria
- `get_filters` — the full list of valid filter values, which is what you call before searching to avoid guessing
- `create_reservation` — books a selected offer and returns a payment link; the booking confirms automatically once payment completes
- `get_reservation_status` — checks a booking using the identifier `create_reservation` returned
- `echo` for connection diagnostics and `test_mcp_authentication` for confirming the token is valid
- An n8n workflow template for a Telegram travel-agent bot, shipped as a demo rather than a production architecture
An access token, sent as an `Authorization: Bearer` header. Nothing runs locally — clients that support remote MCP servers point straight at `https://mcp.winwin.travel/mcp/sse` with the header in their MCP settings. Connecting and searching are free; you pay only when a booking is completed. The README recommends models at an intelligence level equivalent to Claude Sonnet 4.6 or higher, and warns that weaker ones misread tool responses or hallucinate hotel data on complex requests.