Power BI Desktop MCP v2: Automate Semantic Models with AI

In this video I put my Power BI Desktop MCP server v2 to a real test: can an AI agent build a complete semantic model from scratch?

Using Claude Code with local MCP support, I let an LLM drive an 8-phase workflow:

connect to SQL Server
import the Contoso tables
optimize the model
create measures
define UDFs
add age buckets
apply translations
document everything it has done

This is an experiment, not a polished marketing demo. I show where it works well, where it fails, and how easy it is for a powerful MCP to push a model into a corrupted state if you are not careful.

What you will see:

Overview of my MCP server for Power BI Desktop (v2)
How I prompt the LLM to work in clearly defined phases
Running the workflow in Claude Code planning mode with agents
Automatic table import and star schema creation on the Contoso 10k database
Measures, UDFs, age buckets, and a dedicated measures table
Date table configuration, relationships, and basic model hygiene
Spanish translations created via MCP tools
Validating what the LLM claims it did against the actual Power BI model
Token / context usage for a full semantic-model build

Tools used:

Power BI Desktop with my MCP server (Model Context Protocol) v2
SQL Server Contoso sample database (10k row version)
Claude Code with local MCP support
DAX and semantic modeling best-practice documents as extra context

Important:

This MCP is powerful enough to put your model into a corrupted state if misused. Always:

Work on a copy of your PBIX or PBIP
Keep backups / version control
Manually verify relationships, measures, and transformations
Treat the LLM output as a starting point, not ground truth

In this video I put my Power BI Desktop MCP server v2 to a real test: can an AI agent build a complete semantic model from scratch?

Using Claude Code with local MCP support, I let an LLM drive an 8-phase workflow:

connect to SQL Server
import the Contoso tables
optimize the model
create measures
define UDFs
add age buckets
apply translations
document everything it has done

This is an experiment, not a polished marketing demo. I show where it works well, where it fails, and how easy it is for a powerful MCP to push a model into a corrupted state if you are not careful.

What you will see:

Overview of my MCP server for Power BI Desktop (v2)
How I prompt the LLM to work in clearly defined phases
Running the workflow in Claude Code planning mode with agents
Automatic table import and star schema creation on the Contoso 10k database
Measures, UDFs, age buckets, and a dedicated measures table
Date table configuration, relationships, and basic model hygiene
Spanish translations created via MCP tools
Validating what the LLM claims it did against the actual Power BI model
Token / context usage for a full semantic-model build

Tools used:

Power BI Desktop with my MCP server (Model Context Protocol) v2
SQL Server Contoso sample database (10k row version)
Claude Code with local MCP support
DAX and semantic modeling best-practice documents as extra context

Important:

This MCP is powerful enough to put your model into a corrupted state if misused. Always:

Work on a copy of your PBIX or PBIP
Keep backups / version control
Manually verify relationships, measures, and transformations
Treat the LLM output as a starting point, not ground truth