MCP Engine: AI-Native Power BI Semantic Modeling with Production Guardrails

AI can now write DAX, create measures, and restructure your semantic models through natural language. That part is solved. What isn't solved is what happens when it breaks something.

MCP Engine is a local MCP server that bridges AI assistants - Claude, Copilot, ChatGPT, or any MCP-compatible client - with your Power BI semantic models. It gives AI the ability to read, query, and modify your models. And it gives you the ability to control exactly what AI is allowed to do.

Free to start. No credit card. Install in under a minute.

What is MCP?

Model Context Protocol is an open standard that lets AI assistants interact with external tools. Instead of copying DAX into a chat window and pasting results back, MCP lets the AI work directly with your model - reading metadata, executing queries, creating measures, managing relationships - all through a standardized interface.

MCP Engine implements this protocol for Power BI. It connects to the Analysis Services engine running inside Power BI Desktop (or Power BI Service for cloud models) using the same APIs Power BI developers have used for a decade: TOM for metadata, ADOMD.NET for queries. Nothing exotic - just a standardized, AI-accessible wrapper around the tools you already trust.

The speed difference is real

What used to take 15 minutes of clicking through menus - find the table, create a measure, write the DAX, format it, test it - now takes about 5 seconds through a natural language prompt.

Manual workflow vs MCP Engine

But speed without safety is a liability. That's where MCP Engine separates from the pack.

Free tier: everything you need to start

The free tier is not a demo. It covers the core modeling operations that most people need day-to-day:

Natural language DAX - create measures, calculated columns, and calculation groups by describing what you want.
Bulk operations - rename, delete, or update dozens of objects in a single prompt.
Auto translation - translate your model to 40+ languages.
Query execution - run DAX queries and get results directly in the chat.
Model exploration - list tables, columns, measures, relationships, and hierarchies.
Basic performance stats - get a read on query duration and storage engine behavior.
Cross-platform - works on Windows and macOS, in Claude Desktop, Claude Code, VS Code, or any MCP client.

Cross-platform support

Why Pro?

Microsoft released their own Power BI MCP server at Ignite 2025. It's free, and it handles the basics well. If all you need is natural language model editing and bulk operations, their server - or MCP Engine's free tier - will get you there.

But if you're working on production models, or you're a consultant responsible for a client's data infrastructure, the basics aren't enough. You need to know what breaks before it breaks. You need to undo mistakes instantly. You need to control what the AI can and cannot do.

That's what Pro is for.

Rollback and snapshots

Every change MCP Engine makes is automatically recorded. You can preview diffs before applying, undo the last operation, or roll back to any previous state. It's a git-like workflow for your semantic model - without requiring git.

Experiment fearlessly - automatic history and instant rollback

This alone changes how you work with AI. When you know you can undo anything, you stop hesitating and start experimenting.

Test runner

Define expected values for any measure. Run the suite before deployment. If Total Sales should be 1.4M and it suddenly isn't, you'll know before your stakeholders do.

Test runner - catch regressions before deployment

Test results can be exported for CI/CD pipelines, so you can integrate model validation into your existing deployment workflow.

Policy engine

Block risky operations. Enforce naming conventions. Require confirmation before deletes. Your rules, always applied - whether you're the one prompting or someone on your team.

Policy engine - block, allow, or require confirmation for any operation

Policies are declarative. Define them once, and they apply to every session. No more hoping that the AI - or the person talking to it - follows the rules.

Impact analysis

"What breaks if I rename this column?" Instead of finding out in production, MCP Engine traces measure dependencies and shows you the cascading impact before you commit.

Impact analysis - trace dependencies before you change anything

RLS testing

Impersonate roles and users directly from the chat. Validate that row-level security filters actually work without switching contexts in Power BI Desktop.

RLS testing - impersonate roles and validate security without leaving the chat

Data masking

This is the enterprise feature most people don't think about until it's too late. When AI reads your model, it sees your data. MCP Engine's masking uses session-randomized coefficients to hide real values while preserving trends, ratios, and distributions. PII is partially redacted. Your AI assistant works with representative data - never the real numbers.

Data masking - session-randomized coefficients preserve patterns while hiding actual values

Persistent memories

Teach MCP Engine your business logic once: "Our fiscal year starts in July." "We never use SUMX on large tables." "Use DIVIDE() for all ratios." It remembers across sessions, so every interaction starts with the right context.

Memories - persistent business logic across sessions

Token efficiency

Schema and payload optimization cuts token usage in half. Longer sessions before hitting context limits means more complex tasks completed in a single conversation.

Up to 2x longer sessions through token optimization

Performance profiling

Capture full query plans, storage engine stats, and formula engine timings. Understand exactly where your DAX is spending time.

Performance profiling - full query plans and engine timings

Architecture and privacy

MCP Engine runs locally. Zero telemetry. Your model stays in your environment. The server connects to the Analysis Services instance inside Power BI Desktop - the same local engine that's been running your models for years.

Your AI assistant will send prompts and results to its provider (Anthropic, OpenAI, etc.) based on their privacy terms. That's unavoidable with any AI-assisted workflow. But the MCP server itself does not phone home, and with data masking enabled, the AI never sees your real numbers.

Install

One command:

text
claude mcp add mcp-engine -- npx -y mcp-engine

That's it. Works with Claude Code out of the box. For Claude Desktop, VS Code, or other MCP clients, see the getting started guide.

Requirements: Windows 10/11 or macOS. Power BI Desktop with an open PBIX file, or connection to Power BI Service (Premium/Fabric capacity required for cloud models).

Try these prompts

Once connected, try:

"List all tables and their row counts."
"Create a YoY Sales Growth measure using DIVIDE."
"Show all measures that reference the Sales table."
"Run a test suite on my margin measures."
"What would break if I renamed the Customer table?"
"Show VertiPaq stats and recommend columns to reduce model size."

Pricing

Free - core modeling, queries, bulk operations, translations. No limits, no expiry.

Pro - $39/month or $349/year. Rollback, test runner, policy engine, impact analysis, RLS testing, data masking, memories, full performance profiling, and Power BI Service connectivity.

No credit card required to start. The free tier is permanent.

Get started

Build Faster. Break Nothing.

➡️ Website: mcpengine.dev

➡️ Download: GitHub Releases

➡️ Full feature comparison: mcpengine.dev/compare

If you have workflows you want to automate or features you'd like to see, reach out - I'm building this based on what Power BI professionals actually need.