Building Claude Dashboards on Microsoft Fabric with Fabric IQ MCP

Step 1: Create an Azure application

Claude cannot sign in to Fabric on its own. You give it a sign-in by registering one application in Azure, then pointing two custom connectors at that application. Nothing else is needed: no secret, no service principal and no gateway.

In the Azure portal, go to Microsoft Entra ID > App registrations and create a new registration. Then set up two things.

• A redirect URI. Under Authentication, add a platform of type Mobile and desktop applications and set the redirect URI to https://claude.ai/api/mcp/auth_callback. This is what lets Claude receive the sign-in result, and it makes the app a public client, which is why there is no client secret.

• API permissions. User.Read (Microsoft Graph) is there by default. Under API permissions, choose Add a permission, then Power BI Service, then Delegated permissions, and search for and add each of these permissions. None of them required admin consent in our tenant.

API	Permission	What it allows
Microsoft Graph	User.Read	Sign in and read the user profile
Power BI Service	Dataset.Read.All	View all datasets
Power BI Service	Item.Read.All	Read Fabric items
Power BI Service	Item.Execute.All	Make API calls that require execute permission on Fabric items
Power BI Service	SemanticModel.ReadWrite.All	Read and write semantic models
Power BI Service	Workspace.Read.All	View all workspaces

• Copy the Application (client) ID from the app’s overview page. You will paste it into both connectors.

Step 2: Connect Claude to Fabric

You add each Microsoft Fabric MCP server to Claude as a custom connector. Do the steps below twice, once for each MCP server, using the same client ID from Step 1 both times.

Connector name	MCP server URL
Microsoft Fabric IQ	https://fabriciq.svc.cloud.microsoft/v1/mcp/fabriciq
Microsoft Fabric	https://api.fabric.microsoft.com/v1/mcp/core

If your tenant uses private links, the Fabric IQ URL is https://api.fabric.microsoft.com/v1/mcp/fabriciq.

1. Go to Settings, then Connectors. Choose the + Add button.

2. Choose Add custom connector, enter a Name (only a label in your connectors list) and the MCP server URL from the table, then choose Continue.

3. Claude contacts the server and detects that it uses OAuth, so each user will sign in through the server’s OAuth flow before any tool is used. Under Authentication, leave Sign in now selected. Under OAuth client, choose Use your own OAuth client (Claude detects this too) and paste the Application (client) ID you copied in Step 1. Leave the client secret blank, since the permissions are delegated and the app is a public client. Leave Request headers empty and choose Add.

4. Click Connect.

5. This takes you to the Microsoft permissions screen. Review the requested permissions and choose Accept.

6. If you have not already, repeat steps 1 to 5 for the other MCP server. You need both connectors connected: Microsoft Fabric IQ and Microsoft Fabric.

After you connect

From then on, Claude acts as you, so you can only reach what that account can reach, and row-level and object-level security on the model still apply.

The two servers do different jobs.

• Fabric MCP manages items: list workspaces, list items, read an item’s metadata. It has no tool for querying data.
• Fabric IQ MCP queries data. Its tools are DiscoverArtifacts, ResolveFabricItem, GetSemanticModelSchema, ValueSearch, GetReportMetadata and ExecuteQuery, which runs DAX.

With both connected, Claude can find your workspace and model, query it, and build the dashboard from the results, which is what the rest of this post covers.

Step 3: Getting started with Claude Dashboards on Microsoft Fabric

Start a new Claude Dashboard artifact. Go to Artifacts in the sidebar, then choose Dashboard (Beta) from Make something new.

Most of our time went on working out how these servers behave. Paste the block below at the start of a chat, or into a project’s instructions, so Claude can skip that discovery. Replace the placeholders in angle brackets.

I want a Claude Dashboard built from a Microsoft Fabric semantic model.

Workspace: <workspace name or id>
Semantic model: <model name or id>   (or a report ID; it resolves to its model)
Goal: <the questions the dashboard should answer, or "your call">

Optional (leave blank and Claude will ask):
Date role: <which date to use, e.g. order vs ship date; any inactive relationships to use>
Personal data: <e.g. company names only, no personal fields>
Period: <full history, or a range / as-of date>
Currency/units: <if the model mixes them>

How to get the data
- Use the Fabric IQ MCP connector to query the model with DAX. Use the Fabric MCP connector only to find the workspace and model. Don't try the Fabric REST API, Power BI endpoints or login.microsoftonline.com from the sandbox; they are blocked.
- Find the model with DiscoverArtifacts or ResolveFabricItem.
- Schema: EVALUATE COLUMNSTATISTICS(). Measures: EVALUATE INFO.MEASURES(), selecting Name and Expression. Read the definitions and note caveats (cost basis, time intelligence on partial years, which relationship a measure uses).
- COLUMNSTATISTICS shows sample values for every column. Don't query or show personal fields (names, emails, phones, credentials) unless I ask.
- Check every table you plan to use with a multi-row query. A table that returns nothing gets no error. Leave it out and say so on the page.
- Profile before designing: fact row counts, date range, rows per period. Pick the time grain from that.
- Prefer the model's own measures so numbers match existing reports. Aggregate in DAX. Max 4 queries per call. Pass maxRows when a result can exceed 250 rows.

Known quirks of the Fabric IQ connector
- Make every result multi-row. A single-row result comes back as a citation card even when padded, so put totals in long format: UNION(ROW("metric","x","value",…), …).
- Add ADDCOLUMNS(…, "pad", REPT("x", 4000)) so the result comes back as a CSV, not a citation card. Large results are saved to a JSON file: an array of {type, text} items. The CSV is the text that starts with "[Resource from … ]". Parse it with a script and never retype numbers.
- Add a constant tag column to each query ("q", "<dataset id>") and match results by tag, not position. A failed query returns nothing and no error, so the order shifts. Drop the pad and q columns before saving.
- Alias columns with SELECTCOLUMNS. Otherwise headers come back as "Table Column".
- Wrap amounts as ROUND(<measure>, 2) * 1.0, but the parser must still strip a leading apostrophe, "$" and thousands separators. Some measures keep their currency format anyway.
- Don't put FORMAT() or other never-blank expressions in SUMMARIZECOLUMNS; it cross-joins. Use ADDCOLUMNS(SUMMARIZE(…)) or format dates in the parser.
- The dashboard can't query Fabric IQ live. Its rows arrive as an attached resource, and the page receives only the citation. Use file datasets.

Building the dashboard
- Save each result as a CSV and upload each CSV in its own call. Mixed batches are refused.
- Register each as a file dataset. Put its DAX in the dataset description so a refresh re-runs exactly the same query. Then write the page so marks bind to the datasets.
- Before building, reconcile the tables: every breakdown sums to the headline total, counts match across tables, and levels of detail agree (header vs line, fact vs dimension). Put a reconciliation table on the page and show gaps rather than hiding them.
- Add an "About these figures" note: snapshot date, model and workspace, measure caveats, and anything missing and why.
- This is a snapshot. When I say "refresh", re-run the saved queries, compare with the current data, and replace only the datasets that changed. Update every dataset's refresh time.

Before you start, tell me which tables and measures you plan to use, and ask about anything unclear, including any optional field I left blank.

If your model uses different table or measure names, say so in the first block and Claude will adapt the queries.