TensorPass

Use cases

AI trip planning on a travel site, run on the traveler's own AI

How a travel site answers trip questions with each traveler's own AI, which MCP tools it serves, which interface it picks and how booking stays on its pages.

Published on . Updated on

Traffic from AI assistants to US travel sites grew 119% in the year to July 2026. Those visits lasted 67% longer than other visits and had a 42% lower bounce rate. They converted 1% below other traffic, against 47% below a year earlier. AI Connect, the TensorPass SDK a website installs, puts the trip planner on the travel site itself. The traveler's model searches the site's inventory through its tools and answers on the traveler's own account through AI Wallet, so the site pays no inference.

Trip questions travelers put to an AI

In an April 2026 survey of 1,006 US travelers, 54% had used AI to plan a trip. Among the uses they named, 60% research attractions, restaurants and transport, 56% compare prices and 43% build itineraries. Asked about destination sites, 63% would use an AI trip planner there, and 25% are less likely to use a destination site that has no AI features.

Users of Priceline's Penny and of Agoda's assistant converted at about twice the rate of other users in the third quarter of 2026. Users of Expedia's own AI features are twice as likely to book an activity within 7 days. Referrals from AI chat apps stayed well under 1% of Booking.com's room nights over the same quarter.

On a travel site, the questions turn into searches of its own inventory:

Inventory tools for a travel site

The Model Context Protocol (MCP) is the open standard through which a model calls a site's tools. A travel site serves its inventory as an MCP server on its own domain, at a path such as /api/mcp, and the SDK calls it from the traveler's page.

ToolInputWhat it returns
list_destinationsNoneThe places, neighborhoods and property types the site holds, as accepted filter values
search_staysPlace, guests, budget, amenitiesThe matching properties, each with its id and page address
check_availabilityProperty id, arrival date, nights, guestsThe free rooms with their price and terms, and the address of the booking page for those dates
get_propertyProperty idOne property in full: description, amenities, cancellation terms, location
search_activitiesPlace, dateTours and activities with their times, prices and page addresses

The model reads only the tool names, their descriptions and their input schemas. The MCP guide sets the rules: dates explained with their format in the description, every input bounded (guests, nights, a maximum limit), and an empty result that tells the model to relax one filter. The MCP server validates each call's arguments, since anyone reaches it without the page. The docs MCP server of tensorpass.app allows 60 requests a minute per address, and an availability tool that queries a supplier system takes a limit sized to that system.

Which interface suits a trip planner?

The SDK ships five ready made interface modules. Three of them suit a travel site:

Every module ships its labels in English, French, Hebrew, Arabic and simplified Chinese, and lays Hebrew and Arabic out right to left.

Keeping the booking on the site

The tools above read inventory and carry the readOnlyHint annotation, which tells MCP clients that they change nothing. check_availability returns the address of the booking page for the chosen dates as a field of its result. The SDK's answer guard lets a link through when it points to the page's own origin, to a host the site lists, or to an address a tool returned as a field. The traveler follows that link, enters their details and pays in the site's own booking flow.

Booking.com was among the first apps in ChatGPT, where users browse stays and finish the booking on Booking.com. With AI Connect, the booking step of a travel site runs on its own pages in the same way.

Property reviews are third party text. The SDK closes every tool result in an envelope sealed with a random salt drawn for each conversation, and instructions planted in a review stay inside it as data. This is the guard against prompt injection, detailed in the security model.

Next steps for a travel site

On 16 November 2026 the SDK packages of version 0.1.0 are on npm, and AI Wallet waits on the Chrome Web Store's review while a visitor can install it by hand from https://tensorpass.app/wallet. A traveler without AI Wallet gets the planner through a sign in with OpenRouter, or with the TensorPass plan on a site that passes a site key.

Sources

Questions

Who pays for the trip planner's answers?
Each answer runs on the traveler's own ChatGPT, Claude or Gemini account through AI Wallet and counts against that account's plan or credits. The travel site pays for its MCP server and its availability calls.
How much traffic do AI chat apps send to travel sites in 2026?
In the third quarter of 2026, referrals from AI chat apps stayed well under 1% of Booking.com's room nights. Traffic from AI assistants to US travel sites grew 119% in the year to July 2026.
Does the traveler's AI book the room?
The tools in this setup read inventory and change nothing. The check_availability tool returns the address of the site's booking page, and the traveler enters their details and pays there.
What does the site's server learn about the traveler?
The MCP server receives the arguments of each tool call the model makes, such as the place, the dates and the number of guests. The traveler's credentials stay in AI Wallet, and the conversation goes to the provider of the traveler's account.
Have it done in 24 hoursRead the quickstart