For those of you worried about feeding your data to AI: I decided not to go for the MCP route but instead integrated Claude Code in my VS Code which I pointed at my pbip files.
This way you only expose your metadata and framework (dax, m queries, report structure) to AI which I personally feel way more comfortable with
The results in my opinion are very equivalent to the MCP approach. MCP wins at context it can gather from looking at the actual data and being able to use that for validation, but you pay a high price with sending that data into Anthropic's servers.
