Have a conversation with your data using TimeXtender MCP

You may recognize this. You have a question and a dashboard with 40 visuals, but the answer isn't among them. The usual route is to request a change from the BI department. By the time it's delivered, your question is sometimes already outdated.

So you download the data from the dashboard and build your own analysis in Excel.

What if that were a thing of the past? What if you could put your business questions directly to a chatbot such as ChatGPT or Claude?

At the end of January 2026, I got my first hands-on experience with TimeXtender MCP. MCP (Model Context Protocol) forms a bridge between a Large Language Model (LLM) and your semantic data model.

Installation and initial setup

For this test, I set up a separate virtual machine with TimeXtender and the Microsoft Adventure Works demo dataset. That way I could test independently, without any impact on a production environment.

TimeXtender MCP is currently still offered as a preview and is installed separately. The installation itself was straightforward and finished within a few minutes. During configuration, you specify where the server runs and which semantic endpoint(s) are made available.

For this trial setup I chose Claude, Anthropic's AI assistant. MCP can connect to multiple LLMs, but Claude has its own configuration tab, which makes connecting it easy.

Within TimeXtender I created a semantic endpoint, similar to how you set up an endpoint for Power BI, Qlik or Tableau. From the MCP configuration, a Windows service is then set up that makes the data model available to the AI assistant.

The complete setup therefore consists of three components:

• 1. The semantic model for MCP in TimeXtender,
• 2. The MCP service that makes the model available,
• 3. The connection to the AI assistant (Claude).

Why asking questions of a data model is different

The strength of this assistant is that you aren't querying a database, but a data model. The assistant knows the relationships between the entities and understands the business logic it is given, even when it sees it for the first time.

When you ask a question (to Claude in this case), you see it write SQL queries at lightning speed and use them to query the model. No technical knowledge of tables or queries is needed. This makes asking questions accessible, while the answers stay in line with the data model.

The results are presented in a clear HTML view with visualizations, often supplemented with contextual explanation.

Asking questions in practice

I started with a simple question: what was this year's revenue? This was the result.

Then I raised the complexity and asked for a Pareto analysis. What you get back is not just a number, but multiple visualizations with context.

Next, I asked for a regression analysis and tested other statistical challenges. Without having to create a measure or write a query myself, I was presented with a structured answer.

The answer seemed correct, although I didn't verify it down to the last decimal. What stands out most is that the step from question to analysis becomes very small.

Working with multiple models

Something else I wanted to test was whether MCP can handle multiple semantic endpoints. It turns out it can.

For my test, I created a second model and connected it. I then asked questions where Claude could clearly explain the differences between the two models.

This opens up interesting possibilities. You could set up a separate assistant per user group or per business process, each with its own set of data models. The context stays separated, while the interaction remains the same for the user.

What can't it do (yet)?

The possibilities are broad. Still, there are clear limitations.

MCP is not a platform for machine learning or model training. Simulations and training new models are not supported. Analyses are performed on the basis of the existing data model. For specific operations, another MCP, such as the one from Qlik, can be used if needed.

A more important point of attention is security.

In my test environment, Row Level Security (RLS) was not yet applied. The assistant does not automatically recognize that my Claude user and my TimeXtender user are the same person. As a result, the entire model can be queried, regardless of any access restrictions.

For use in a production environment, this is an essential point that needs to be set up properly.

The importance of context

You can refine the model, and with it the assistant, by adding extra context. In TimeXtender you do this easily by including clear descriptions for measures and fields.

For example: in the description of the measure 'revenue' you record that revenue is always shown in euros. The assistant takes this context into account when interpreting questions and generating answers.

The better the definitions and context are captured in the model, the more consistent the results will be. A well-maintained semantic model thus becomes an important prerequisite for the responsible use of AI within BI.

Conclusion and outlook

My first experience with the preview version of TimeXtender MCP shows how quickly the interaction between AI and BI is developing. The generated output is clear and well designed, but more importantly, the answers align with the underlying data model.

At the same time, there are still points of attention, particularly around security and user context. For use in a production environment, these are essential prerequisites.

What this development mainly shows is that the needs of the business are changing. There is demand for direct answers to ad-hoc questions, without first having to build a new dashboard or an additional visual. MCP offers a way to answer those questions within the existing business logic of the model.

For BI teams, this doesn't mean their role gets smaller, but different. Less time will go to one-off ad-hoc requests and more to carefully setting up and maintaining semantic models. Adding clear definitions and context therefore becomes even more important.

The focus shifts from reporting to carefully designing and managing the data model.

Ultimately, it is not the assistant but the quality of the model that determines how valuable the answers are. And those who have their model in order can truly have a conversation with their data.