Dr. Estera Kot

FabCon Europe 2026 in 18 demos · Part 3 of 18 · MCP servers

Fabric IQ MCP server: query a semantic model from your coding agent

A remote, read-only MCP server lets GitHub Copilot CLI or any MCP client find a semantic model and run DAX as you. Your agent writes the DAX.

By Dr. Estera Kot, Microsoft MVP, CTO at Clouds on Mars, and former Microsoft Principal Product Manager for the Apache Spark runtime in Fabric. Published .

The Fabric IQ MCP server is a remote Model Context Protocol (MCP) server that connects AI agents and MCP clients to Power BI reports and semantic models. It exposes six read-only tools that find content, read metadata, search values, and run DAX queries as the signed-in user. Microsoft Learn lists it as generally available. In FabCon Europe 2026 in 18 demos, it is the door for developer tools rather than people.

A semantic model is now one prompt away from any coding agent. You write no connector code and create no service account.

In short

  1. Fabric IQ MCP is GA: a remote, read-only MCP server with six tools that find a Power BI report or semantic model and run DAX as the signed-in user.
  2. It has no answering tool. Your coding agent picks the tools, writes the DAX, and explains the result.
  3. It enforces your permissions, RLS, and OLS. The gate to watch is consent, because users can grant the two Power BI scopes themselves by default.
Real calls to the Fabric IQ MCP server from a terminal: six read-only tools, the synthetic Oud RnD model found by name, and one DAX query with Oud Saffron S-7 on top at 68.0. 23 seconds, no sound, recorded Oct 21, 2026.

Same demo in Polish and German. Step-by-step runbook: D03.

What the Fabric IQ MCP server is and how it works

Clients connect over Streamable HTTP and sign in with delegated OAuth 2.0 through Microsoft Entra ID. Every client needs two Power BI Service permissions: Item.Read.All and Item.Execute.All.

The server applies the caller's existing Fabric permissions. You need no workspace role and no Build permission on the model. Service principals and app-only sign-in are not supported, and there is no administrator impersonation path.

The path of one Fabric IQ MCP request Four boxes from left to right. Your coding agent, such as GitHub Copilot CLI, writes the DAX. Microsoft Entra ID signs the user in with delegated Item.Read.All and Item.Execute.All, and users can consent by default. The Fabric IQ MCP server offers six read-only tools and no answering tool. The semantic model applies the AI setting for read-only users and RLS and OLS. Result rows return to the agent, 250 by default. Your coding agent Microsoft Entra ID Fabric IQ MCP Semantic model GitHub Copilot CLI or any MCP client Delegated sign-in Item.Read.All Item.Execute.All 6 read-only tools Runs DAX as the user AI setting for read-only users RLS and OLS apply Writes the DAX Users can consent by default No answering tool Always enforced Result rows return to the agent: 250 by default, more as an embedded CSV Yours Operated by Microsoft The path of one Fabric IQ MCP request From top to bottom: your coding agent writes the DAX, Microsoft Entra ID signs the user in with delegated permissions that users can consent to by default, the Fabric IQ MCP server offers six read-only tools, and the semantic model applies the AI setting, RLS, and OLS. Result rows return to the agent. Your coding agent Microsoft Entra ID Fabric IQ MCP Semantic model GitHub Copilot CLI or any MCP client Delegated sign-in: Item.Read.All Item.Execute.All 6 read-only tools Runs DAX as the user AI setting for read-only users, then RLS and OLS Writes the DAX Users can consent by default No answering tool Always enforced The right-hand line: result rows return to the agent, 250 by default. Yours Operated by Microsoft
Figure 1. One Fabric IQ MCP request. Microsoft runs the server and the model, and both enforce your permissions. Your side writes the DAX and decides where the rows go next.

The semantic model applies its own AI setting before any rows leave. Read-only users pass only if the model allows AI, see part 2, the semantic model AI setting.

What the demo shows

The 23-second clip makes real calls from a terminal, through a small script named fabric_iq_mcp.py. The runbook runs the same three steps through fabdemo.py.

The 68.0 matches the expected answer that the recap computes from the synthetic trial data.

Limits and catches

The catch: the two Power BI scopes need no admin consent by default. Any user who can read a model can connect an MCP client of their choice, and the rows land in that client's language model. If that breaks your policy, restrict user consent or require the admin approval workflow.

The server does not answer questions. Learn: “Fabric IQ MCP doesn't expose a natural-language answering tool.” Figure 2 shows who does each step of one question.

  1. Find the item Server tool DiscoverArtifacts or ResolveFabricItem
  2. Read the structure Server tool GetReportMetadata, GetSemanticModelSchema
  3. Find exact values Server tool ValueSearch, when available
  4. Write the DAX Your agent, no tool Quality depends on your model's metadata
  5. Run the query Server tool ExecuteQuery, one model per call
  6. Explain the result Your agent, no tool The rows reach your agent's model
Figure 2. The six steps of one question, in Learn's order. The server runs four. Your agent does the two that decide whether the answer is right.

One tenant setting has a similar name and a different job. Users can use the Power BI Model Context Protocol server endpoints (preview) controls the Power BI Model Authoring MCP experience. Learn says it does not enable or disable Fabric IQ MCP.

Scope is narrow. The server covers reports and semantic models only, not dashboards, paginated reports, apps, ontologies, or data agents. For agents, see part 12, the Fabric data agent. ExecuteQuery returns 250 rows unless you set maxRows.

It needs a tenant home region that supports all Fabric workloads. Power BI-only regions and sovereign clouds are out. RLS can give two users different answers to the same query.

How to connect and query the Fabric IQ MCP server

You need the Oud RnD model from the demo script. Runbook D03 also lists the PowerBIMCP tenant setting, turned on by settings on. Per Learn, that setting governs a different server.

python3 fabdemo.py iq-mcp tools
python3 fabdemo.py iq-mcp discover "Oud RnD"
python3 fabdemo.py iq-mcp dax "Oud RnD" 'EVALUATE SUMMARIZECOLUMNS(formulas[formula_name], "Preference %", [Avg first choice %]) ORDER BY [Preference %] DESC'

Expected result: six read-only tools, the model found by name, and the average first-choice share per formula with Oud Saffron S-7 on top at 68. Time and cost: about 3 minutes. Queries use the capacity, which must be Active.

To query from a coding agent, add Learn's configuration for GitHub Copilot CLI to ~/.copilot/mcp-config.json, then run /mcp show FabricIQ and sign in. Learn recommends the Fabric IQ skill from Skills for Fabric alongside it.

{
  "mcpServers": {
    "FabricIQ": {
      "type": "http",
      "url": "https://fabriciq.svc.cloud.microsoft/v1/mcp/fabriciq",
      "tools": ["*"]
    }
  }
}

Do not add an Authorization header. Learn warns that it overrides interactive OAuth and token refresh.

What to do on Monday

On Monday: Pilot Fabric IQ MCP on one workspace with one developer.

For the server that writes, read part 4, the Fabric Core MCP server first.

Key facts at a glance

Checked on Microsoft Learn on November 3, 2026.

IDFactSource
K1Status: generally available.Get started with Fabric IQ MCP
K2Endpoint: fabriciq.svc.cloud.microsoft/v1/mcp/fabriciq. With private links: api.fabric.microsoft.com/v1/mcp/fabriciq.Get started with Fabric IQ MCP
K3Six read-only tools, from DiscoverArtifacts to ExecuteQuery.Get started with Fabric IQ MCP
K4Delegated Item.Read.All and Item.Execute.All. No admin consent by default.Get started with Fabric IQ MCP
K5No service principal, no app-only sign-in, no admin impersonation.Get started with Fabric IQ MCP
K6ExecuteQuery: one model per call, 250 rows by default.Get started with Fabric IQ MCP
K7Pin the tool contract with X-Variants: Fabric.Routing.FabricIQ.V1.Get started with Fabric IQ MCP

Frequently asked questions

Is the Fabric IQ MCP server generally available?

Yes. Microsoft Learn states that Fabric IQ MCP is generally available, as checked on November 3, 2026. It needs a tenant home region that supports all Fabric workloads.

What tools does the Fabric IQ MCP server have?

Six read-only tools: DiscoverArtifacts, ResolveFabricItem, GetReportMetadata, GetSemanticModelSchema, ValueSearch, and ExecuteQuery. There is no natural-language answering tool, so the calling agent writes the DAX and explains the result.

Can a service principal use the Fabric IQ MCP server?

No. Learn states that service-principal and application-only authentication are not supported. The server uses delegated OAuth 2.0 and runs every call under the signed-in user's Power BI permissions.

Does the Power BI MCP tenant setting turn Fabric IQ MCP on or off?

No. The setting Users can use the Power BI Model Context Protocol server endpoints (preview) controls the Power BI Model Authoring MCP experience. Learn states that it does not enable or disable Fabric IQ MCP.

What is the difference between Fabric IQ MCP and Fabric Core MCP?

Fabric IQ MCP is GA and read-only, with six tools for querying Power BI reports and semantic models. Fabric Core MCP manages workspaces, items, roles, and OneLake settings. Its Learn tools reference listed 43 tools on November 3, 2026, and 25 of them write.

About the author

Dr. Estera Kot is CTO at Clouds on Mars, a Microsoft and Databricks partner for data and AI in Europe and the Middle East. At Microsoft she was Principal Product Manager for the Apache Spark runtime in Azure Synapse Analytics and Microsoft Fabric, and she created the Native Execution Engine. She is a Microsoft MVP and holds a Ph.D. in Computer Science from the Warsaw University of Technology. She presented this series as the FabCon Europe 2026 recap in Gdynia on October 21, 2026, and online on October 23, 2026. Slides, recordings, and runbooks are in the public recap repository.

Sources

Microsoft Learn, checked on November 3, 2026, then the recap materials.