Microsoft Fabric now includes MCP (Model Context Protocol) servers and a command-line interface that let AI agents operate directly on Power BI resources, data workspaces, and pipelines. In a Fabric Insider episode published June 27, Reza Rad of RADACAD interviewed Hasan Abo-Shally, Product Manager for Fabric MCP Servers and the Fabric CLI, and demonstrated agents creating lakehouses, provisioning workspaces, and automating pipelines purely through natural language prompts.
Two MCP Modes: Local and Remote
Fabric offers two MCP server options with different risk profiles. The Local MCP is open-source and read-only, designed for development, code generation, and safe experimentation. It lets agents read Fabric metadata and generate CLI scripts, but humans review and approve before anything changes.
The Remote MCP is a cloud-hosted, authenticated endpoint that lets agents perform real actions inside a Microsoft Entra ID-protected tenant. It supports dry-run modes, allowing teams to simulate agent actions before committing them. According to the HubSite365 report, remote execution demands careful identity and access controls, with tenant permissions and role-based controls governing what agents can do.
Agent-Agnostic by Design
The integration works with GitHub Copilot, Claude Code, and ChatGPT through the standardized MCP protocol. Abo-Shally emphasized that model capability influences prompt interpretation and error handling, but the underlying protocol standardizes how agents connect to Fabric regardless of which LLM powers them. More capable models produce accurate automation faster but also enable higher-risk actions if governance is weak, requiring organizations to evaluate reliability alongside cost and access policies.
The Fabric CLI as Execution Surface
Live demonstrations showed the Fabric CLI as the execution layer for agent outputs. Agents can bulk-manage semantic models, update Power BI reports, and access OneLake data programmatically. The CLI produces agent-generated scripts that are staged for human review, combining automation speed with manual oversight.
Platform-Native vs. Bolt-On Agents
The move reflects a broader enterprise adoption pattern: data platforms that integrate agents natively reduce friction compared to bolting agents on top of existing APIs. Rather than requiring developers to reverse-engineer API endpoints and build custom tooling, MCP-compliant servers provide a typed API that lets agents understand and control Fabric resources in a predictable format.
Microsoft’s roadmap includes tighter tooling and community contributions to open-source MCP server components, according to the Fabric Insider episode. The open question is whether organizations will adopt the governance processes necessary to give agents production access to data infrastructure, or whether dry-run mode becomes the permanent default.