Microsoft Fabric IQ now feeds Copilot by default, and opens an MCP endpoint for outside agents

A Microsoft 365 Copilot that answers a business question now reads from the same Power BI numbers the analysts trust, by default. Microsoft said at its FabCon and SQLCon conference in Barcelona on 29 September that Fabric IQ, the context layer inside Microsoft Fabric, is generally available inside Copilot Chat and Cowork, with no extra AI token cost, and that a Model Context Protocol server for Fabric IQ is open to any compatible agent.

The New Stack, which covered the keynote, reports that Fabric IQ combines data from Microsoft OneLake, metrics from Power BI semantic models, and operational rules from ontologies into one layer that both people and agents can query. Microsoft Fabric CTO Amir Netz described agents as forgetful by nature and said the point of the layer is to tell them "you are now working for Microsoft, you are now working for Wells Fargo, you are now working for Emirates" before they do anything else.

The MCP endpoint, and what it exposes

Microsoft's own documentation for the Fabric IQ MCP server, now listed as generally available, describes it as a remote MCP server with a read-only toolset that lets an agent find Power BI content, inspect report and semantic model metadata, locate values, and run DAX queries. It respects the signed-in user's existing Fabric permissions, including Power BI row-level and object-level security, so connecting through MCP does not grant access to anything the caller could not already read. The server does not expose a natural-language answering tool, so the calling agent still composes the queries and explains the results.

Microsoft listed GitHub Copilot CLI among the first clients, with a dedicated Fabric IQ skill. Any MCP client that supports streamable HTTP can connect without a separate Fabric integration.

Power BI gets an agentic app builder

The second announcement is agentic app creation in Microsoft Power BI Desktop. Microsoft says Power BI has more than 425,000 customers. The new experience lets a user start from a trusted semantic model, describe an application in natural language, preview and edit it, and publish it to Fabric. The apps accept input, write data back and preserve shared state, with each app getting a SQL database in Fabric, authentication and security through what Microsoft calls Fabric Apps. They are open to any Power BI user, not restricted to Fabric capacity customers, and will reach preview in the coming weeks. Backends can be built with the open source Rayfin SDK and CLI, which now adds TypeScript backend functions, PostgreSQL support and a Secret Store.

Mirrored data, now with the permissions attached

Microsoft also said at the keynote that mirrored security roles will reach public preview in the coming weeks. These bring permissions along with the data when an external source is mirrored into OneLake. The example on stage was Snowflake: access roles will show up in Fabric with the same members and the same table permissions. The company also announced a two-way integration with Salesforce Data Cloud 360, general availability of mirroring from Google BigQuery, and new connections with Atlan, dbt and ClickHouse.

For a team building an agent that reads from a Microsoft 365 estate, the practical change is that business definitions are now in one place with a standard protocol in front of them, instead of being rebuilt per project. The caveat is that most of this depends on being inside Microsoft's stack already: a company that does not use Fabric, OneLake and Power BI gets no benefit from the Fabric IQ layer, and the mirrored-security guarantees apply only to sources Microsoft has written a mirroring connector for.

Source

Microsoft Fabric is where AI agents learn how the business works by Frederic Lardinois, The New Stack, 30 September 2026. Primary source: FabCon and SQLCon 2026 in Barcelona: Building the data foundation for Microsoft Copilot and agents, Microsoft Azure Blog, 29 September 2026, and Get started with the Fabric IQ MCP server, Microsoft Learn.