Your Power BI model now takes orders from an LLM

What it does

The Power BI Modeling MCP Server is a local bridge that exposes Power BI semantic models—whether in Desktop, Fabric, or project files—to any MCP-compatible AI client. Instead of manually editing relationships, measures, or columns, you describe the change in plain text and the agent manipulates the model directly. It handles bulk operations like mass renames, translations, and query refactors with transaction support, and can execute or validate DAX queries against the connected dataset.

The interesting bit

Microsoft is treating the semantic model as just another codebase that an agent can plan against, edit via TMDL, and commit back—complete with an “elicitation” protocol that forces the LLM to ask for human confirmation before its first modification or query. It is essentially an admission that BI modeling has become too tedious to do by hand at scale.

Key highlights

• Supports natural-language creation and modification of tables, columns, measures, and relationships.
• Executes bulk modeling operations—renames, translations, security rules—across hundreds of objects at once.
• Integrates with Power BI Desktop, Fabric workspaces, and PBIP/TMDL project files for agentic development workflows.
• Requires user confirmation before the first model change or DAX query unless explicitly configured otherwise.
• Ships primarily as a VS Code extension but can be registered in any MCP client.

Caveats

• The project is in Public Preview and Microsoft warns that the implementation may change significantly before General Availability.
• The README explicitly warns that LLMs can hallucinate changes and may leak sensitive metadata or data into chat logs; backups and caution are advised.
• It is strictly a modeling tool—it cannot touch report pages, visual layouts, or diagram arrangements.

Verdict

Power BI developers and data engineers drowning in repetitive semantic-model maintenance should try this; if you do not work with Power BI semantic models, there is nothing here for you.