Insights / Research

Research

Testing Claude Code on a published Power BI semantic model

We had Claude Code read, query and edit a published Power BI model through Microsoft's MCP server. Figures matched; the hosted server stopped at sign-in.

Upside Analytics R&D5 min read

At Upside Analytics R&D, we tested whether Claude Code can work on a Microsoft Power BI semantic model that is already published, with nothing done by hand in Power BI. Through the local Power BI Authoring MCP server, it could read the model, query it and make a test edit. The Microsoft-hosted version of the server stopped at sign-in in our setup.

The question

In Power BI, the semantic model is the layer between the data and the reports: tables, relationships and the business calculations reports use. Those calculations are measures, named formulas written in DAX, such as "total assets".

The Model Context Protocol (MCP) is an open standard that lets an AI assistant use external tools. Microsoft's Power BI Authoring MCP server is one such tool: it lets an assistant read a semantic model, run queries against it and change it.

Upside Analytics R&D had already used Claude with this server to build models from Power BI Project (PBIP) files on disk. This time we wanted to know whether the same approach works on a model that is already published, and whether the newer hosted version of the server changes anything.

Two versions of the server

According to Microsoft's documentation, the server comes in two deployments that do the same job in different places:

Can connect toLocal (generally available)Hosted (preview)
Semantic models in a Fabric workspaceYesYes
Models open in Power BI DesktopYesNo
PBIP and TMDL files on diskYesNo

The local server runs on your own machine. The hosted server runs on Microsoft's infrastructure, so there is nothing to install, but it can only reach models that are already published.

How we tested it

On October 28, 2026, we used Claude Code with VERSION_CLAUDE as the AI client and the local Power BI Authoring MCP server VERSION_MCP, and pointed it at our Treasury demo semantic model, which was already published. We had PERMISSION permission on the model. All work went through the MCP server; nobody opened Power BI Desktop or edited the PBIP files during the test.

What we found

TaskResult
Connect to the published model (local server)Worked
List measuresAll 34, with descriptions
Return and check five figuresAll 5 matched the reference
Add a test measure in a transaction, check it, roll backWorked; model still had 34 measures
Sign in to the hosted serverCouldn’t sign in

The five figures we checked

Claude wrote and ran the queries itself. Each result matched our reference value at the precision shown in the reference:

IndicatorValue returned
Total assets$49,019M
Total liabilities−$48,373M
Net position$646M
Overnight funding share57.1%
Net position on September 30$764M

The demo model uses synthetic treasury data.

Editing the model

Through the local server, Claude created a temporary test measure inside a transaction, so the change was not committed to the model. A regular DAX query could not see the uncommitted measure. We checked its formula instead with a query-scoped DEFINE MEASURE, which defines the measure for that one query only. We then rolled back the transaction and confirmed that the model still had 34 measures.

The hosted server

When we tried to sign in to the hosted server, Claude Code could not complete the sign-in. Microsoft's documentation explains a likely cause: some MCP clients depend on dynamic OAuth client registration, which Microsoft Entra ID does not support. Microsoft suggests using the local server with those clients, or registering a Microsoft Entra app yourself. The hosted endpoint also requires a Fabric administrator to enable a tenant setting.

We did not try the Entra app route, because the local server already covered what we needed. Our result applies to our Claude Code setup; other clients may behave differently.

What it means for your team

  • Start with the local server. It reaches published models, Power BI Desktop and PBIP files, and it worked with Claude Code without extra tenant configuration.
  • Check permissions first. Microsoft's documentation says Write permission on the semantic model is needed to change it. With only Build permission, the assistant can run queries but cannot edit.
  • Decide what data may leave your environment. Model metadata, such as table and measure names, and query results are sent to whichever AI provider your client uses. In regulated environments, confirm that this is approved before you connect a real model.
  • Keep models in PBIP format under Git. Every change the assistant makes then shows up as a diff you can review or revert. Microsoft recommends the same, and suggests backing up the model before you start.
  • Test against known answers. Before trusting new figures, ask for a few you have already verified and compare.

Limitations and next steps

  • We tested one demo model with one AI client, Claude Code.
  • We checked five figures, not every measure in the model.
  • We did not get the hosted server working, so we cannot compare the two versions in practice. A next step is to register a Microsoft Entra app and repeat the test.
  • This post covers the model test on October 28. We tested report building (Claude creating report pages and reviewing its own screenshots) on the same project on October 14; that test is not covered here.

Sources

Want to talk about this?

We build governed reporting and AI for BI for financial services teams.

[email protected]