Edit

Use AI agents with OneLake through MCP

The Model Context Protocol (MCP) is an open standard that lets AI agents connect to external tools and data sources. The Fabric MCP server includes OneLake tools that give AI agents direct access to your OneLake data — browsing workspaces, reading files, discovering table schemas, and managing directories — all through natural-language conversation.

Fabric items use open formats to store their data in OneLake — from lakehouses to mirrored databases, and even KQL databases and semantic models with OneLake availability enabled. The OneLake MCP tools let AI agents explore all of these through a single set of commands.

Prerequisites

Install the Fabric MCP server

The simplest way to get started is with the Fabric MCP server VS Code extension. Install the extension and the OneLake tools are available automatically.

For manual configuration or use outside VS Code, see the setup instructions in the Fabric MCP Server README.

Authentication and permissions

The OneLake tools use your existing Azure identity and Fabric permissions. Your AI agent can only access workspaces and items you already have permission to view. You don't need any additional roles or permissions beyond your normal workspace access.

Sign in with the Azure CLI before using the tools:

az login

OneLake MCP tool capabilities

The tools cover three areas of OneLake:

Category What you can do OneLake APIs used
Workspace and item discovery List workspaces, list items within a workspace, create new items (lakehouses, notebooks, SQL databases, and more). OneLake data plane API and Fabric REST API
File and directory operations Browse, read, write, upload, download, and delete files and directories. Both DFS and Blob Storage endpoints are available. OneLake file system APIs
Table operations Get table API configuration, list namespaces (schemas), list tables, and retrieve full table definitions with column names, types, and metadata. OneLake table APIs

The tools include 19 commands. Most commands accept friendly names as well as GUIDs — for example, you can reference an item as SalesLakehouse.lakehouse instead of a GUID.

For the full command reference with parameters and example output, see the OneLake tools README on GitHub.

Example: explore a lakehouse

After you install the Fabric MCP server, ask your AI agent to explore your data. For example:

"List the tables in my Sales lakehouse and tell me about the schema of each one."

The agent uses the OneLake MCP tools to locate the workspace, find the lakehouse, discover the table namespaces, and retrieve the full table definitions — without writing any code or opening the Fabric portal.

Other tasks you can ask your AI agent to complete:

  • Inventory a workspace — "Scan my workspace and tell me what items are there, how big they are, and what tables they contain."
  • Read and write files — "Upload this CSV to the Files folder in my lakehouse" or "Read the config file in the Monitoring folder."
  • Document a mirrored database — "Describe the tables and storage structure of my mirrored database."
  • Create items — "Create a new lakehouse called StagingData in my Analytics workspace."

Limitations

  • The item create command currently requires a GUID-based workspace ID. Your AI agent handles this automatically by calling onelake workspace list to resolve the workspace name to an ID.
  • The OneLake MCP tools don't include shortcut management commands. To list or create shortcuts, use the OneLake shortcuts REST API.