Building with AI in Mendix: Daryna's experiences developing EmergoQuest with AI

What happens when an AI assistant doesn't just think along about a Mendix app, but can also read and change the model directly?

While developing EmergoQuest, a quiz app for E-mergo's 25th anniversary, our colleague Daryna experimented with Claude Code and mxcli. She discovered where AI can speed up the development process, but also why a passed technical check doesn't mean everything works as intended. In this article, she shares her experiences and key takeaways.

What happens when an AI assistant doesn't just give advice about your Mendix app, but can also explore and modify the model on its own? I was able to test that question in practice while developing EmergoQuest, a versatile quiz app I built for E-mergo's 25th anniversary.

For this project, I combined Mendix Studio Pro with Claude Code and mxcli. Thanks to mxcli, the AI agent could explore the Mendix model and prepare and carry out changes. I set the direction, made the decisions and checked the result.

That sped up development considerably, but it also taught me an important lesson: “Check passed” doesn't mean everything works.

Using AI as an assistant and sparring partner during development is no longer new. With mxcli, I went one step further: the agent could explore the existing Mendix model itself and make changes to it directly.

My workflow consisted of three steps: check → execute → verify.

First I checked exactly what would change, then the change was executed, and afterwards I verified whether the result was actually correct.

Over the course of the project, this produced more than sixty scripts for model changes, checks, translations and documentation.

The biggest gain from AI was in repetitive work.

EmergoQuest includes different types of quiz questions, themes, badges, scores and an admin environment. I could leave many recurring changes, checks, translations and documentation tasks to the AI agent.

That left me more time for complex microflows, thinking through the logic, making decisions during development and the creative side of the application.

So not less thinking, but more room to focus on the parts where my knowledge and experience as a developer make the difference.

One of the most interesting moments came when I asked the agent to run a full check of the model after its changes.

During that check, the agent found a problem with an association around player scores: the ownership was set incorrectly.

A small setting with big consequences. Players could, in effect, ‘steal’ each other's scores. The agent itself indicated that the earlier change had not been carried out correctly and suggested possible fixes.

So the AI first helped build part of the model, and then, while checking its own work, discovered that it contained a mistake.

Interestingly, a similar problem appeared later with new associations. That made it clear that AI is perfectly capable of checking its own work and finding important errors, but that this doesn't guarantee the same mistake will be prevented in the future.

That's why I tried to turn these kinds of findings into explicit rules and additional checks wherever possible.

Not everything can be assessed by checking the model technically. Some problems only became visible when I actually used the application.

A component could be technically correct while its behaviour in the app still didn't match what I had in mind. That's why testing the application in Studio Pro remained a fixed part of my workflow.

A green check means that a specific technical check has passed. Not that the application automatically does what you expect it to.

I deliberately kept security and other sensitive parts in my own hands. I also made the most important changes to microflows manually.

In addition, I kept a PROJECT_STATE document recording important decisions, existing components and consciously accepted risks. CLAUDE.md contained the rules and agreements the agent had to follow.

Because the more an AI agent can do on its own, the more important it becomes to set clear boundaries.

At first, I was mainly curious how much faster I could develop with AI.

After this project, I find a different question much more interesting: which work can I leave to AI, so I can focus on the parts where my experience and choices as a developer really matter?

AI can explore the model, prepare changes, run checks, keep documentation up to date and help map out risks.

But in the end, it's still the developer who understands the context, makes the decisions and judges whether the result actually works in practice.

EmergoQuest started as a quiz for E-mergo's anniversary. For me, it also became a valuable experience with a different way of Mendix development, in which AI doesn't just advise but actively contributes to the development process.

Would you like to build a use case with AI in Mendix too? At E-mergo, we're happy to help you discover and apply the possibilities of AI-enabled development. We combine our hands-on experience with AI and our Mendix expertise to arrive at smart, workable solutions together.

Curious what AI can do for your Mendix project? Feel free to get in touch.