Co-authored by

What Is MCP and Why Does It Matter for Power BI?

Quick answer: The Model Context Protocol (MCP) is an open standard that lets AI assistants connect to external tools through a structured, secure interface — and for Power BI, it means Claude can read your semantic model, write DAX, and manage measures without you copying and pasting anything between windows.

MCP defines a client-server architecture where an AI application (the host) connects to one or more MCP servers that expose specific tools. Each server declares what it can do — list tables, execute a DAX query, create a measure — and the AI model calls those tools as needed during a conversation.

For Power BI professionals, this solves a real workflow problem. Before MCP, using Claude with Power BI meant manually copying model metadata into a chat window, pasting DAX back and forth, and hoping the AI had enough context to give accurate suggestions. If you are specifically interested in how Claude handles DAX formula generation, we cover that workflow in depth separately. With MCP, Claude connects directly to your open Power BI Desktop file (or a Fabric semantic model) and works with the live model structure.

There are now two official Microsoft MCP servers and several community alternatives. This guide covers how to set up the most practical options with Claude Desktop and Claude Code.

What You Need Before Starting

Quick answer: You need Power BI Desktop with a .pbip file open, an MCP server installed (Microsoft's official one or a community alternative), and either Claude Desktop or Claude Code as your AI client.

Here is the full prerequisites list:

Enable Power BI Preview Features

Before connecting any MCP server, enable these preview features in Power BI Desktop under File > Options > Preview features:

• Power BI Project (.pbip) save format — saves your model as editable files rather than a binary .pbix
• Store semantic model using TMDL format — enables text-based model definitions that MCP servers can read and modify
• Enhanced metadata format for reports — improves compatibility with external tooling

Save your file in .pbip format via File > Save As before proceeding. The MCP server connects to the Analysis Services instance that Power BI Desktop runs locally — your .pbip file must be open.

Setting Up the Power BI MCP Server

Quick answer: There are three main options — Microsoft's official Modeling MCP server (best for VS Code + Copilot), Microsoft's Remote MCP server (cloud-only, Fabric required), and community servers like MCP Engine (best for Claude Desktop). Pick based on whether you work locally or in Fabric.

Option 1: Microsoft's Official Modeling MCP Server

The Power BI Modeling MCP Server is Microsoft's first-party solution. It is in public preview as of April 2026 and provides 20+ tool categories covering tables, columns, measures, relationships, DAX queries, calculation groups, RLS roles, partitions, hierarchies, and more.

Installation via VS Code (recommended by Microsoft):

1. Install VS Code with the GitHub Copilot and GitHub Copilot Chat extensions
2. Search the VS Code Extensions marketplace for "Power BI Modeling MCP Server" by Microsoft (or visit the marketplace listing)
3. Install the extension
4. Enable "MCP servers in Copilot" in your GitHub Copilot settings at github.com (enterprise admins must enable this at the org level)

Manual installation for Claude Desktop:

Since the VS Code extension bundles the server executable, you can extract it for use with Claude:

1. Download the VSIX package from the VS Code marketplace
2. Rename the .vsix file to .zip and extract it
3. Locate the server executable at extension/server/powerbi-modeling-mcp.exe
4. Copy the entire extension/server/ folder to a stable location like C:\MCPServers\PowerBIModelingMCP\

The Claude Desktop configuration for this server is covered in the next section.

Key CLI arguments:

Option 2: Microsoft's Remote MCP Server (Fabric)

The Remote Power BI MCP Server is a hosted endpoint at https://api.fabric.microsoft.com/v1/mcp/powerbi that connects to cloud-hosted Fabric semantic models. It uses Copilot's DAX generation engine and authenticates with your Microsoft Entra ID credentials.

Requirements:

• Fabric capacity (F2 or higher, ~$262/month minimum)
• Admin must enable the tenant setting: "Users can use the Power BI Model Context Protocol server endpoint (preview)"
• Build permissions on the target semantic model

This option is read-and-query focused — it generates and executes DAX queries against existing models but does not modify model structure. It is currently optimized for VS Code with GitHub Copilot, not Claude Desktop.

Option 3: MCP Engine (Community — Best for Claude)

MCP Engine is a community-built MCP server specifically designed to work well with Claude Desktop and Claude Code. It connects to Power BI Desktop's local Analysis Services instance and provides a streamlined tool set.

Free tier includes 9 read-focused tools: model exploration, DAX query execution, natural language querying, bulk operations, and auto-translation across 40+ languages.

Pro tier ($349/year) adds 6 additional tools: rollback and snapshots, test runner for measure validation, policy engine, impact analysis, RLS/OLS testing, and data masking.

Installation for Claude Code:

claude mcp add mcp-engine -- npx -y mcp-engine@latest

Installation for Claude Desktop: Add to your config file (see next section):

{
  "mcpServers": {
    "mcp-engine": {
      "command": "npx",
      "args": ["-y", "mcp-engine@latest"]
    }
  }
}

MCP Engine runs locally, collects zero telemetry, and supports both Power BI Desktop and Power BI Service via XMLA endpoints.

Configuring Claude Desktop

Quick answer: Add the MCP server to your claude_desktop_config.json file, restart Claude Desktop completely, and look for the tools indicator in the chat input area.

Config File Location

You can also access this file from Claude Desktop: click the Claude menu (system menu bar, not the in-app settings), select Settings, navigate to the Developer tab, and click Edit Config.

Configuration for the Microsoft Modeling MCP Server

{
  "mcpServers": {
    "powerbi-modeling-mcp": {
      "command": "C:\\MCPServers\\PowerBIModelingMCP\\extension\\server\\powerbi-modeling-mcp.exe",
      "args": ["--start"],
      "env": {}
    }
  }
}

For read-only mode (recommended when exploring an unfamiliar model):

{
  "mcpServers": {
    "powerbi-modeling-mcp": {
      "command": "C:\\MCPServers\\PowerBIModelingMCP\\extension\\server\\powerbi-modeling-mcp.exe",
      "args": ["--start", "--readonly"],
      "env": {}
    }
  }
}

Use double backslashes (\\) for Windows paths in JSON. A single backslash will cause a silent parse failure that disables all MCP servers.

Configuration for MCP Engine

{
  "mcpServers": {
    "mcp-engine": {
      "command": "npx",
      "args": ["-y", "mcp-engine@latest"]
    }
  }
}

After Saving the Config

1. Quit Claude Desktop completely — closing the window is not enough. On macOS, use Cmd+Q. On Windows, right-click the system tray icon and select Quit
2. Relaunch Claude Desktop
3. Look for the tools indicator — a small hammer icon should appear in the bottom-right of the chat input area. Click it to see the available Power BI tools
4. Open Power BI Desktop with your .pbip file — the MCP server needs an active Power BI instance to connect to

If the tools indicator does not appear, check the logs at ~/Library/Logs/Claude/mcp*.log (macOS) or %APPDATA%\Claude\logs\mcp*.log (Windows).

Configuring Claude Code

For Claude Code (the terminal-based client), use the claude mcp add command:

# Microsoft Modeling MCP Server (Windows)
claude mcp add powerbi-modeling -- "C:\MCPServers\PowerBIModelingMCP\extension\server\powerbi-modeling-mcp.exe" --start
 
# MCP Engine (cross-platform)
claude mcp add mcp-engine -- npx -y mcp-engine@latest

Verify the connection with /mcp inside a Claude Code session.

Your First AI-Assisted Semantic Model Session

Quick answer: Start by asking Claude to connect to your local Power BI Desktop instance and describe the model, then move to specific tasks like writing measures, validating DAX, or documenting your model structure.

Here is a realistic workflow once the MCP server is connected:

Step 1: Connect and Explore

Start a conversation in Claude Desktop and ask it to connect to your local Power BI model:

"Connect to my local Power BI Desktop instance and tell me what tables and relationships are in the model."

Claude will use the MCP server's connection and table tools to list your model structure. This is a good sanity check — if Claude returns your actual tables, the connection is working.

Step 2: Understand the Model

Ask Claude to audit what already exists:

"How many measures do I have? List them by table, and flag any that reference columns that don't exist."

This is where MCP starts paying for itself. Manually checking measure dependencies across a large model is tedious. Claude can read every measure's DAX expression and cross-reference column references in seconds.

Step 3: Write a Measure

Ask Claude to create a new measure:

"Create a measure called 'Revenue YoY %' in the Sales table that calculates year-over-year revenue growth using SAMEPERIODLASTYEAR. The revenue column is Sales[Revenue] and the date table is Calendar[Date]."

Claude will generate the DAX and — if the MCP server is not in read-only mode — offer to add it directly to your model. On the Microsoft Modeling MCP server, you will get a confirmation prompt before any write operation is applied.

Step 4: Validate and Test

Ask Claude to run a DAX query to verify the measure works:

"Execute a DAX query that returns Revenue YoY % by year for the last 3 years."

The MCP server's DAX query tools will execute the query against your local model and return the results. Claude can then tell you whether the numbers look reasonable.

Step 5: Batch Operations

One of the most practical uses is bulk refactoring:

"Rename all measures that start with 'Calc_' to remove the prefix. Show me the list of changes first."

Claude will generate the plan, show you the before/after, and apply the changes on your confirmation.

What Claude Can and Cannot Do via MCP

Quick answer: Claude can read model metadata, write and validate DAX, manage measures and relationships, and execute queries — but it cannot create report visuals, modify dashboard layouts, or access data that exceeds your existing Power BI permissions.

The distinction matters: MCP gives Claude access to the semantic model layer — the tables, measures, relationships, and DAX logic. The report layer — visuals, layouts, formatting — remains manual work in Power BI Desktop.

Enterprise Considerations

Quick answer: The local MCP server processes everything on your machine and no data leaves your network through the MCP connection itself — but the AI conversation (including any model metadata or query results discussed) is sent to Anthropic's API, so enterprise teams need to evaluate this against their data governance policies.

Data Flow and Privacy

Understanding exactly what goes where is critical for regulated environments:

1. Local MCP server to Power BI Desktop: Communication happens through operating system pipes (stdio) on the same machine. No network exposure. The MCP server connects to Power BI's local Analysis Services instance on a dynamic localhost port.

2. MCP server to Claude: The server returns metadata and query results as structured JSON through stdio. This data becomes part of the conversation context.

3. Claude conversation to Anthropic: Here is where data leaves your machine. The conversation — including any model metadata, DAX expressions, and query results that Claude has discussed — is sent to Anthropic's API for processing. Anthropic does not use business customer data for model training (per their usage policy), but the data does transit their infrastructure.

For the Remote MCP server (Fabric), the data flow includes Microsoft's cloud infrastructure as well. Queries are processed through the Fabric API endpoint and Copilot's DAX generation engine.

Practical Recommendations for GCC Enterprises

For organizations in the UAE, Saudi Arabia, or wider Gulf region where data residency matters:

• Start with read-only mode (--readonly flag) to understand what metadata is exposed before enabling write operations
• Use local MCP servers rather than the Remote MCP server to avoid sending data through additional cloud endpoints
• Review what gets discussed — if your model contains sensitive column names, table names, or measure logic that constitutes trade secrets, be selective about what you ask Claude to analyze
• Evaluate Anthropic's data processing terms against your organization's data protection requirements (UAE PDPL, DIFC Data Protection Law, Saudi PDPL)
• Do not use --skipconfirmation in production environments — the confirmation prompt is a governance control

As an official Claude partner, Beyond The Analytics helps enterprises onboard Claude safely — from MCP configuration to governance policies. Book a consultation to get your team started.

Row-Level Security Implications

The MCP server respects your existing RLS configuration. If your Power BI Desktop model has RLS roles defined, the server accesses data as the currently connected user. However, during local development in Power BI Desktop, RLS is only enforced when you explicitly use "View as Role." The MCP server connecting to your local instance will see all data by default — just as you do when editing the model.

For testing RLS through MCP, the MCP Engine Pro tier offers an RLS impersonation tool that lets you validate security roles without switching contexts in Power BI Desktop.

How This Differs from Power BI Copilot

Quick answer: Power BI Copilot is Microsoft's built-in AI for report consumers (asking questions about data in natural language), while Claude + MCP is for model developers (building and modifying semantic models programmatically). They serve different audiences and different workflows.

If you have already been evaluating Power BI Copilot for enterprise readiness, it is worth understanding where Claude + MCP fits:

These are complementary, not competing. A development workflow might use Claude + MCP to build and test the model, then publish it to Fabric where business users interact with Copilot. For a full side-by-side evaluation, see our Power BI Copilot vs Claude AI comparison.

Frequently Asked Questions

Q: Does the MCP server send my Power BI data to Anthropic?

The MCP server itself does not send data anywhere — it runs locally and communicates through stdio pipes. However, any model metadata, DAX expressions, or query results that Claude discusses in the conversation are part of the prompt sent to Anthropic's API. The MCP connection is local; the AI conversation is not. Evaluate Anthropic's commercial terms against your data governance requirements.

Q: Can I use Claude + MCP with Power BI Service without Fabric?

Not with the Microsoft Remote MCP server — that requires Fabric capacity (F2 or higher). For local development, you can use the Modeling MCP server or MCP Engine with Power BI Desktop at no additional Microsoft cost beyond your existing license. MCP Engine's Pro tier also supports XMLA endpoints for connecting to Power BI Service or Fabric, but XMLA access itself requires Premium or Fabric capacity.

Q: Which Claude model works best for DAX and semantic modeling?

As of April 2026, Claude Opus 4 and Claude Sonnet 4 both work well for DAX generation and model reasoning. Opus 4 has stronger performance on complex multi-step modeling tasks (like planning a full star schema refactor), while Sonnet 4 is faster and cheaper for routine measure creation and validation. Both support MCP tool use. If you are using Claude Code, you can switch models per session based on task complexity.

Q: Does this replace Tabular Editor or DAX Studio?

No. Claude + MCP is a conversational interface to your model, not a replacement for dedicated modeling tools. Tabular Editor still offers finer control over TMDL, best-practice rules, and CI/CD pipelines. DAX Studio remains the standard for query profiling and server timings. Think of Claude + MCP as a fast way to explore, prototype, and batch-modify — then validate with your established tooling. Tabular Editor has also published guidance on how AI agents and their tool complement each other.

Q: Is this setup production-ready?

The Microsoft Modeling MCP server is in public preview as of April 2026. Microsoft's documentation explicitly states that the implementation "may significantly change prior to General Availability." Use it for development and prototyping. For production model management, continue using your established CI/CD pipeline (TMDL + Git + Tabular Editor) and treat Claude + MCP as an accelerator within that workflow, not a replacement for it.