Inspiration
MCP - a model context protocol, and how it can be very useful in building beautiful applications even when some applications are not able to modernize. REST communication and how to leverage MCP server and client. Legacy applications and traditional web interfaces like Google's Online Boutique demo app needs to be modernized, but in some cases without touching the legacy code. Clicking through shopping categories, using those dropdown filters, opening product pages in new tabs to compare products, is not a very user-friendly experience. For example, adding a chat with AI layer UI on top of the boutique app could make shopping feel more natural?.
What it does
Instead of clicking through, users can just chat with the system like they're texting a friend who works at the store. You can say stuff like "I need running shoes under $100" or "show me wireless headphones with good reviews" and it actually understands what you mean. The system has three main parts working together:
A coordinator agent that figures out what you're asking for A catalog expert that handles all the product search and recommendations A cart manager that deals with adding/removing items and keeping track of your order
It pulls real product data from the original boutique services, so pricing and inventory are always up to date.
You can have back-and-forth conversations too, say for example: if you ask for "something professional" it might show you business clothes, then you can say "more casual" and it adjusts.
The whole thing runs on top of the existing Online Boutique microservices without changing any of the code. So all the original web interface still works.
How we built it
Used Model Context Protocol (MCP) as a bridge. Built an MCP server that translates between AI agent requests and the boutique's gRPC APIs. This meant all the existing online boutique services (product catalog, cart service, checkout, etc.) completely untouched. Used GKE, Python, Gemini, FastAPI and Streamlit. Separated out tasks with separate agents Used Terraform for scripting infrastructure.
Challenges we ran into
RPC stubs creation and methods details. Figuring out exact request to be sent is a challenge and multiple trail and errors went into this. Without terraform help for infrastructure as a code, its a challenge to build and deploy all docker containers at will. Definitely need it. Debugging containers, why the container is not starting, from where the image is getting pulled etc. is little challenging on GKE platform.
Accomplishments that we're proud of
No changes to Online Boutique services. Shopping experience can be more human in natural language, its fun to interact with AI. Used 3 ai agents for this and learned how to communicate between them.
What we learned
MCP, MCP and MCP - this is the one most important take away - it helps in different ways to make beautiful applications without touching the original applications.
What's next for AI-Powered Personal Shopping Concierge
Sky is the limit. Enhance it with voice ai and/or voice interface, enhanced personalization experience with capturing and learning from individual users interactions.
Log in or sign up for Devpost to join the conversation.