This summary is produced by the author, and not by AI.
AI agents are now capable to help you build semantic models. They perform their changes in code, but you can't just stare at TMDL and JSON to understand what's going on; that's like driving a car with mud all over the windshield. Especially in the early phases of developing a model, you need to be able to closely follow that your intent is adhered to and unfounded assumptions don't make their way from the agent's training data into your semantic model. Agents are fundamentally a new actor you need to command and understand, and that requires better tools. We're deeply conscious of this and realize co-developing your semantic model in Tabular Editor with an agent on the side can feel like a hassle, almost an awkward dance to avoid stepping on each other's toes. If you're already using our CLI, Maxim Anatsko's SemanticOps MCP, or Microsoft's Power BI Modeling MCP server to let an agent work on your model, you should be able to review its work in Tabular Editor as it happens.
That's why we're working hard on better AI integration, and already want to lift the veil from a new feature releasing soon: the MCP server, planned for September 23. An MCP (Model Context Protocol) server makes an application's tools available to an agent; TE3's connects it to the model you have open, so you can work together in the same session.
When agents are moving on your model, your work shifts to reviewing; identify what has changed, what was the previous state, and how this change impacts other things, and ultimately decide whether to reject, keep, or further adjust the change. Code diffs are a valuable but rudimentary way to look at those changes in TMDL/JSON; they don't show how it fits into the whole. You need to see the object structure, relationships, and properties. In Tabular Editor, you can select an affected object and inspect its properties and expression, then adjust or undo the change. Connecting the agent to that same session also gives it access to your unsaved edits, so it can work with the model as you currently see it.
Here's how you give your agent access to the Tabular Editor session (this is in a nutshell what an MCP server does; open a door in your application the agent can walk through, with access limited to the tools allowed by your permissions):
The examples here use Claude Code. The agent uses its own AI configuration; you don't need to configure an AI provider or API key in Tabular Editor for the MCP connection. Read our guide on how to get started with agentic development if this is all new to you.
In the first recording, we changed the expression for MTD Gross Sales from DATESMTD to DATESYTD without saving. The agent reads the open model and identifies that the expression now calculates year-to-date sales, despite the measure's month-to-date name, and expresses its concern over this.
If you permit it, the agent can also edit your model, for instance set the appropriate Sort By Column property for month and weekday labels in a Date table. In the recording below, you can see the changed property in TE3, undo the agent's changes with Ctrl + Z, and redo them. The model remains unsaved in this example, so you can review the assignments before saving.
Agents will very soon be able to work on the same model you have open in Tabular Editor 3, including visibility on unsaved changes made by you. You can inspect what changed and decide what to keep, adjust or undo. See you next week in the release blog for Tabular Editor 3.27.0!
Take your semantic models further with Tabular Editor.
Give Tabular Editor a spin