We designed a travel agent–style conversational flow that leads with discovery: questions about interests, group size, pace, and trip constraints. Based on responses, the system generates personalised suggestions presented in an interactive widget, allowing users to pick and refine activities. When a user is ready, the model synthesises those selections into a time-aware itinerary widget with clear timing and next steps, including a call to action.

The diagram above shows how the components fit together in a ChatGPT app. The UI widgets are hosted on a frontend server and then the "backend" is hosted separately on an MCP server. Some MCP tools on the server have references to the UI widgets on the frontend and know how to inject the data into those components, in our case a list of recommended activites returned for a user.

This approach deliberately balances free-form natural language interaction with structured UI to handle complex decisions — a hybrid pattern that tests where chat adds value and where traditional UI is more effective.

OpenAI suggests building experiences that make sense as conversational workflows that extend ChatGPT's capabilities, rather than just mirroring whats on your existing app or website. These widgets are fairly minimal with only a primary and maybe even a secondary action at most.

A small disclaimer: in the prototype video the model returns a timetable that has a couple of missing activities from those that were selected in the previous step. This is a bug with how the SDK is working at the time of writing and recording and will be fixed and re-recorded once resloved.