Most AI agents fail at business questions for a simple reason: they cannot reach the governed numbers. The Fabric IQ MCP server changes that for Power BI. It is a remote Model Context Protocol (MCP) endpoint that lets an AI agent find reports and semantic models, read their structure and run DAX queries, using the permissions of the person who signed in. Microsoft lists it as generally available.

This guide is for BI leaders and architects deciding whether to put agents in front of their semantic models. You will see what the server does, which tools it exposes, how security works, and the limits to plan around before you roll it out.

Why Agents Need Governed Access to BI

Without a governed route, teams copy data into prompts or build one-off connectors, and every copy escapes row-level security and the model owner's definitions of revenue, margin or churn. An agent that queries the semantic model itself inherits those definitions, so an answer about "last quarter's churn" uses the same measure as the executive dashboard.

MCP is the open standard that makes this repeatable. Instead of a custom integration per tool, any MCP-compatible client can call the same server. If you are weighing MCP against direct APIs or command-line tools, our MCP vs CLI vs API decision framework covers the trade-offs.

"An agent that queries the semantic model inherits your definitions. An agent that queries a copy invents its own."

What the Fabric IQ MCP Server Is

Fabric IQ MCP is an MCP server, not a REST API, so clients follow the MCP lifecycle and call tools. It is read-only and it has no built-in natural-language answering tool. The calling agent chooses which tools to use, writes the DAX and explains the result. That keeps the intelligence in your agent and the data access in Fabric.

It works with Power BI reports and semantic models only. A report does not need Fabric or Premium capacity to be reachable, but the tenant's home region must support all Fabric workloads. Sovereign clouds and Power BI-only regions are not supported.

The Six Tools an Agent Gets

Microsoft documents six tools. Clients should still call the server's tools/list at runtime rather than hard-coding names, because the list and its schemas are authoritative.

A typical agent run is short: discover the model, read its schema, write a DAX query, run it, then explain the result. Because the schema comes from the model itself, well-named tables, measures and descriptions directly improve answer quality. This is the same discipline we recommend in our Power BI consulting work.

Microsoft also recommends pairing the server with the Fabric IQ skill, part of its Skills for Fabric package. The skill teaches the agent how to identify Power BI content, interpret report and model context, and sequence the six tools to answer a business question. The server works without it, but your client then has to supply that orchestration and those guardrails itself.

Anatomy of an Agent Session

The sequence below is an illustration, not a recorded transcript. Imagine an analyst asks an agent: "Which region had the biggest drop in on-time delivery last quarter?" A well-behaved agent works through the tools in a predictable order.

Notice that the server never decides what the answer means. It supplies discovery, structure and execution. That division of labour is why the quality of your model matters more than the choice of agent: the agent can only reuse measures and names that exist.

Making Semantic Models Agent-Ready

Microsoft documents the server, not model hygiene, so the following are our recommendations from Power BI delivery work. They cost little and they directly reduce wrong or ambiguous answers.

Security: Existing Permissions Still Apply

Authentication is delegated OAuth 2.0 through Microsoft Entra ID, with the Power BI Service permissions Item.Read.All and Item.Execute.All. The server grants no extra access: the agent sees what the signed-in user can see, and row-level and object-level security still apply. Service principals and application-only authentication are not supported, and there is no administrator-impersonation path, so there is no background identity with broad rights.

Two consequences matter for governance. First, agent answers are only as private as the user's own access. Microsoft notes that a user does not need a workspace role or Build permission on a semantic model to query it through the server, so review everyone who can open your sensitive models, not just the people who build on them. Second, the two permissions do not require administrator consent by default, which means users can usually consent themselves unless your tenant restricts user consent or routes requests through an approval workflow. Involve your identity team early to decide which of those you want. Our data governance consulting team helps set these rules.

Limits to Plan Around

Several constraints shape what you can promise your business users:

Operational Details Admins Should Know

A few details from Microsoft's documentation affect how you deploy and support the server. If your organisation uses private links, the endpoint host differs from the standard one: use https://api.fabric.microsoft.com/v1/mcp/fabriciq instead, and confirm the address with your network team. Clients should also send the X-Variants header with the value Fabric.Routing.FabricIQ.V1 on every MCP request, including initialize, tools/list and tools/call, which pins the tool contract version. The v1 in the endpoint path is just the address, not the contract version.

Governance has one trap. The Power BI tenant setting for the Power BI MCP server endpoints (preview) does not control Fabric IQ MCP, so do not assume that switch governs agent access. Use Entra consent policy, Fabric permissions and your client approval rules instead.

Finally, watch query load. Every question an agent asks becomes a DAX query against the semantic model, and Power BI throttling applies. If the model sits on Fabric or Premium capacity, that load lands on the capacity, so add agent traffic to your capacity monitoring from day one; our Fabric observability guide shows how.

Where It Fits in Your Analytics Stack

Think of Fabric IQ MCP as the question-answering path, and keep authoring elsewhere. Microsoft's separate Power BI Authoring MCP server (formerly the Modeling MCP server) is for building and changing semantic models, and our guide to the Power BI Authoring (formerly Modeling) MCP covers that side. A common pattern is to use the authoring server to build and document a model, then use Fabric IQ so that everyone else can ask it questions. Fabric IQ MCP is the right tool when a developer, analyst or agent needs to ask a trusted model a question.

Good early use cases are an analyst assistant inside a developer tool, a support agent that checks a metric before answering, and a testing agent that validates measures against expected values. Avoid exposing it to large audiences until you have measured the query load it creates.

A Practical Rollout Plan

Numlytics helps teams prepare semantic models for AI, from naming and security to AI and machine learning services. Speak with a certified consultant to plan a pilot.