After showing our demo to Johan and Lise, this week’s topic was MCP: Model Context Protocol. We aren’t adding it to the Engestofte application yet, but I can see where it could become useful later.
Say you have an AI assistant and want to ask it what’s happening in your application. It needs a way to look that up. MCP gives it a standard way to connect to the application’s tools and data. Instead of copying details into a chat, you can let the assistant look up an event or check availability when you ask. That’s the connection described in the official MCP documentation.
The useful part is that the assistant can work with current information. It doesn’t have to guess what is booked or rely on a document someone pasted yesterday. The connection gives it a way to ask the application.
What that could mean for Engestofte#
Imagine Johan asking: “What is still missing for Saturday’s wedding?”
An assistant connected to our application could retrieve the event and its outstanding tasks, then give him an overview. Or he could ask which holiday houses are available for a particular weekend, and the assistant could call the same availability logic that the dashboard uses.
We would have to build and expose those tools. MCP provides the connection; our API still needs to supply the answers and enforce the rules. The model shouldn’t decide whether a house is available when the booking system can check it.
That is what makes the idea interesting to me. Johan could ask a question in his own words and get information that currently takes several clicks to gather. The existing dashboard would still be there, but an assistant could offer another way into the same workflow.
Useful later, rather than another feature now#
For our current project, we’re keeping the focus on the application we’ve already built and getting its workflow ready for the exam. MCP would add another integration to develop and test, and we don’t need it to make the calendar, inquiries, or allergy overview useful.
It also brings us back to the data questions from the previous security post. We’d still need to talk to Johan about what an assistant could access and what would be sent to a model provider. A standard connection doesn’t settle those choices.
If we explore it later, I’d start with a few read-only tools: retrieve an event, list outstanding tasks, and check availability. That would let us find out whether asking the application questions actually saves Johan time before giving an assistant permission to change anything.
For now, MCP is a possible next step. The appeal is being able to connect an assistant to work the application already knows how to do, while keeping the business rules in the system we’ve built.
This is part eleven of my posts on this semester’s AI-driven applications elective.