Alexa+ Travel Planner

Talk to a multi-agent travel planner through an Alexa+ style assistant and MCP

Comment

Inspiration

I'd just finished building a multi-agent travel planner, and it only worked through a Streamlit page. Then I read that Alexa+ integrations run on MCP, and I wondered: what if I could just tell the planner what I want instead of filling in a form? That's what this project is.

What it does

You say something like "Plan a 5 night Goa trip under 22000 rupees, I like beaches and food." A simulated Alexa+ assistant works out which tool to call on my MCP server, the planner builds the trip, and the assistant reads back a short answer.

The server has four tools:

  • plan_trip makes a plan and gives you a short summary and a trip ID
  • get_trip_details tells you the transport, hotel and activities
  • check_budget checks the plan against your budget and lists anything that's still wrong
  • adjust_trip changes the budget or preferences and plans again

Trips are saved in SQLite, so the trip ID still works after I restart the server.

How I built it

The planner was already done before the hackathon. It has transport, hotel and activity agents running in parallel with LangGraph, and a validator written in plain Python that checks the budget and sends only the broken part back to be fixed.

During the hackathon I built:

  • an MCP server in Python over Streamable HTTP (it reports protocol version 2025-11-25)
  • short, speakable replies, with the full details kept in a separate tool
  • a simulated Alexa+ web app: a FastAPI backend that acts as the MCP client and uses Gemini to pick the tool, plus a web page where you can talk or type and hear the answer. You can see every tool call in the chat
  • SQLite storage for trips
  • an MIT licence and a README that says what was there before and what I added

Challenges we ran into

  • The MCP quickstart code used FastMCP, but the newest mcp version renamed it, so my server wouldn't even start. I fixed it by pinning mcp<2.
  • I didn't have Node.js, so I couldn't use the MCP Inspector. I wrote a small Python client to test with instead.
  • My first test trip failed with "Max turns (10) exceeded". The planner worked fine with a different trip, so I used that one to check that the MCP part was working.
  • The async errors were hidden behind a vague "unhandled errors in a TaskGroup" message. I had to unwrap them before I could see what was really going wrong.

Accomplishments that we're proud of

  • My planner now sits behind a standard MCP server and you can control it by voice.
  • The whole flow works end to end: speak, tool call, plan, spoken answer, then follow-up questions about the same trip.
  • Trips survive a restart.
  • You can see the tool calls in the demo, so it's clear what the assistant is actually doing.

What we learned

  • How MCP works in practice: how tools are described, how Streamable HTTP works, and how an LLM picks a tool.
  • Voice answers need to be short. A long itinerary is fine on a screen but awful to listen to.
  • How to read async errors without panicking.
  • How to keep track of what I built before a hackathon versus during it.

What's next for Alexa+ Travel Planner

  • Try it with a real Alexa+ setup instead of a simulation
  • Swap the mock data for live places, weather and hotels
  • Make adjust_trip re-run only the agent that needs changing
  • Add Amazon Bedrock and deploy the server on AWS
  • Show an itinerary card in the chat with MCP Apps

Built With

Share this project:

Updates

Submission history