agenticdirect.io

projects / travel-concierge

Travel concierge

An agent that spots a flight booked without a hotel, decides whether and how to make contact, then helps the traveller choose a hotel through a short conversation instead of a search grid. A proof of concept: still in progress, not production-ready.

Agent Travel Astro MIT ● In progress

Try it

Try it in the browser

Chat with an agent in your browser, in test mode, with your own test credentials.

Open playground

The idea

Once a flight is booked, the traveller is usually left with a hotel search grid, or an email nobody opens. A travel concierge works the other way round. It already knows where and when you’re going, so it can get in touch at the right moment, ask a couple of questions about the trip you want, and come back with a short, reasoned list.

Hotels are the first thing it helps with here. The same approach carries on through the trip: transfers, insurance, things to do and a table for the last night.

What it does

Try it in the playground →

The demo follows one fictional customer, Sarah Chen, who has just booked two return seats from London Heathrow to the Maldives for 15–22 August with no hotel.

  1. Reads the booking and the customer. Route, dates, passengers, how far ahead she booked, her loyalty tier, past stays and how often she has added a hotel before.
  2. Decides whether to make contact. It spots the missing hotel, checks inventory for the dates, estimates an attach probability (73% here) and picks the channel and timing: an in-app message within 30 minutes of the booking.
  3. Plans the follow-ups. Travel insurance once a hotel is booked, a resort transfer two weeks before departure, and check-in tips two days out.
  4. Has the conversation. On Sarah’s phone, the concierge offers help with a stay. She can say she’s sorted, ask for a reminder later, or ask it to find somewhere.
  5. Asks about budget, then priorities. A slider shows the hotel cost and the full trip total as she moves it; she can also just type a number. Then she picks what matters: house reef, overwater villa, adults-only, spa, family-friendly or fine dining.
  6. Comes back with three hotels. Picked from her budget tier and ranked by how many of her priorities each one meets, with a best match flagged. Each has a detail view with highlights and a map.
  7. Completes the trip. A payment summary (flights already paid, hotel due today), then an itinerary with buttons for transfers, activities and a farewell dinner.

A “behind the scenes” log runs alongside, showing each decision and the tokens used.

How the playground is built

The concierge’s replies are written live by Claude Haiku through this site’s server, on a fixed prompt that knows Sarah’s booking. Short markers in its replies tell the page when to show the budget slider, the hotels or the confirmation. Everything else is scripted: the analysis, the decision, the follow-up plan, and which card appears when.

The hotels are a fixed set of nine real Maldives properties across three price tiers, at illustrative prices. They are not live availability.

Where it is

Next steps:

  • Swap the fixed hotel list for live results from the Hotelbeds MCP server, behind the server-side proxy Hotelbeds needs for signing.
  • Replace the marker approach with native tool calling, so the agent can research hotels mid-conversation instead of the page deciding what to show.
  • Drive the contact decision from real booking and profile data rather than a scripted analysis.

Limitations

A proof of concept, in progress. The customer, trip, prices and payment are fictional, the analysis is scripted and the hotel list is fixed. Nothing can be booked or paid for.