WanderMind — Agentic Travel Concierge for Alexa+

• 

  WanderMind

Inspiration

Travel planning is broken: a "simple" weekend trip means juggling 10+ tabs — flights, hotels, calendars, reviews, bookings. Voice assistants can answer questions, but they can't do the work. We asked: what if Alexa+ could act like a real travel agent — one that knows you, coordinates every service autonomously, and actually books the trip?

What it does

WanderMind turns a single sentence ("Plan me a long weekend in Tokyo next month — keep it under $2,500") into a fully booked trip:

• Agentic orchestration — the agent autonomously calls four MCP services (flights, hotels, calendar, experiences), reasoning over results to compose complete, budget-aware itineraries.
• Cross-session memory — it recalls your seat preference, stay style, and past trips (e.g. "you loved the Osaka food tour") and applies them weeks later without being asked.
• Purchasing — one tap confirms flight + hotel + experiences, charged to a saved card via Alexa+ purchasing.
• Rich media — itineraries render as swipeable cards and carousels, not walls of text.

How we built it

• Alexa+ / LLM agent loop as the orchestrator, with tool-use over MCP servers (flights.search, hotels.search, calendar.check, experiences.search).
• MCP Apps expose each travel service as a composable tool server the agent discovers and calls autonomously.
• Agent Skills encapsulate reusable behaviors: budget guardrails, itinerary composition, preference matching.
• Persistent memory store keeps traveler profiles across sessions and updates after every booking.
• React + TypeScript front-end renders the conversational UI, tool-call trace, media cards, and purchase flow.

Challenges

• Orchestration reliability: chaining dependent calls (flights before hotel location filtering) required explicit planning steps, not naive parallel tool calls.
• Memory design: deciding what to remember (preferences vs. one-off requests) took iteration — we settled on explicit memory writes after each completed booking.
• Trust in purchasing: users won't let an agent spend money blindly, so we designed a clear confirm step with full price transparency before any charge.

What we learned

Agentic UX is a trust gradient: show the tool calls, show the reasoning, and keep the human in control of the money. The winning pattern is autonomy with visibility.

Built With

• agent-skills
• agentic-ai
• lexa-plus
• llm
• mcp
• mcp-apps
• react
• tailwindcss
• travel
• typescript
• voice-assistant