Power BI Modeling MCP Server: let an agent edit your model, but on your terms

An AI agent can now make changes inside your Power BI semantic model. Microsoft has released the Power BI Modeling MCP Server in public preview: a local server that lets an agent such as GitHub Copilot in VS Code create and change tables, columns, measures, relationships and more, using plain language.

This is not another chat box on top of a report. It is write access to the model itself, which is why it is useful, and why you should set it up carefully.

What the server is

MCP, the Model Context Protocol, is an open standard for giving AI agents tools. The Power BI Modeling MCP Server implements it for semantic models and runs locally on your machine. Microsoft lists it as a preview feature in the November 2025 Power BI update, and ships it as a Visual Studio Code extension.

You point it at a model in one of three places:

• An open file in Power BI Desktop.
• A semantic model in a Fabric workspace.
• A Power BI Project (PBIP) folder, by opening the TMDL definition folder.

The README lists a long set of tools behind the scenes. They cover tables, columns, measures, relationships, partitions, hierarchies, calculation groups, security roles, perspectives, Power Query parameters, translations and DAX user-defined functions. There are also batch tools for bulk changes, transaction tools (begin, commit, rollback) and tools to run and validate DAX queries.

Where it earns its place

The examples Microsoft gives are telling. They are not "build my model from scratch". They are the chores most of us postpone:

• Analyse naming conventions and bulk rename objects to match.
• Add descriptions to every table, column and measure.
• Generate a translation of the model, for example into French.
• Refactor similar measures into a calculation group.
• Move hardcoded data source details into semantic model parameters.
• Benchmark a DAX query against two versions of a model.
• Generate Markdown documentation of the model, including relationships and row-level filters.

After two decades in data, I know how much time disappears into renaming, describing and documenting. An agent doing the first pass, with you reviewing it, is a real gain.

Safety first: work on a copy

Microsoft's own warning is direct. The underlying model may produce unexpected or inaccurate results, which can lead to unintended changes, so always create a backup of your model before you start. The README also warns that the model's data or metadata can end up in logs or chat responses, so be careful when you share chat sessions.

There are some guardrails built in. The server asks for your approval before the first change to a model and before the first query against it. You can start it with a --readonly option that blocks all writes. There is also a --skipconfirmation option, and Microsoft says to use it only when you are confident about what will happen and have backups. I would leave it off.

Permissions matter too. The agent acts with your Fabric permissions, and Microsoft notes that an autonomous or misconfigured client can perform destructive actions. Use least privilege, and do not point it at a production workspace while you learn.

Source control is the real safety net

The setup I would recommend is PBIP with TMDL, under Git. TMDL stores the model as a folder of readable text files, with separate files per table, role, perspective and culture. That gives you clean diffs.

The workflow then looks like any other code change:

1. Save the model as a Power BI Project in TMDL format and commit it.
2. Create a branch for the agent's work.
3. Connect the server to the PBIP definition folder, or to a copy of the model.
4. Ask for one focused change, such as descriptions for one table.
5. Read the diff before you commit. Reject what you do not understand.

A backup protects you from disaster. A diff catches the small, plausible mistakes, like a "tidied up" measure that now returns a different number.

The limits to know

This is a preview, and the README says tools and implementation may change significantly before general availability. Other points worth knowing:

• It only performs modeling operations. It cannot change report pages or the diagram layout.
• It follows the same rules as modeling operations from External Tools in Power BI Desktop, so the existing External Tools limitations apply.
• Connecting to a model in a Fabric workspace may not work in your tenant yet, because the client ID used for authentication is still rolling out.
• Data can leave Microsoft's compliance boundaries. The server may share data with third-party clients and language models, and you are responsible for checking that this fits your organisational and regulatory requirements.
• The software may collect usage data and send it to Microsoft.

If you work with Danish or EU customers, discuss that point before connecting an agent to a model with real customer data.

What to do next

If you want to try it this week, keep it small:

1. Install VS Code, GitHub Copilot and the Power BI Modeling MCP extension.
2. Pick a test model, saved as PBIP with TMDL and committed to Git.
3. Start with read-only tasks, like generating documentation or analysing naming conventions.
4. Move on to one write task, such as descriptions, and review the diff.
5. Agree with your team or customer which models are allowed and which are off limits.

Takeaway

The Power BI Modeling MCP Server gives agents real hands on your semantic model. Used on a copy, with TMDL in Git and a human reviewing every diff, it can take the drudgery out of renaming, documenting and translating. Used directly on production, it is a risk you do not need.

Which modeling chore would you hand to an agent first, and which model would you keep it well away from?

Sources

• Power BI Modeling MCP Server README at launch (GitHub, Microsoft)
• Previous monthly updates to Power BI Desktop and the Power BI service, November 2025 update (Microsoft Learn)
• Power BI Desktop project semantic model folder, TMDL format (Microsoft Learn)
• External tools in Power BI Desktop, data modeling operations (Microsoft Learn)