Microsoft Fabric MCP Server

Microsoft Fabric MCP Server

A Python MCP server that lets you manage Microsoft Fabric through natural language in Claude Code or Claude Desktop. 77+ tools covering workspaces, lakehouses, warehouses, SQL, DAX, semantic models, notebooks, pipelines, OneLake, and Microsoft Graph.

Inspired by: https://github.com/Augustab/microsoft_fabric_mcp/tree/main

Quick Start

git clone https://github.com/Jasuni69/ms-core-mcp.git
cd ms-core-mcp
az login
python setup.py

The setup script checks prerequisites, installs dependencies, and configures your Claude client. It safely merges into existing config files without affecting other MCP servers.

Then restart VS Code or Claude Desktop and ask: "List my Fabric workspaces"

Prerequisites

• Python 3.12+ - Download
• uv package manager - Install
• Azure CLI - Install
• ODBC Driver 18 (optional, for SQL tools) - Download
• Access to a Microsoft Fabric workspace with Contributor or Admin role

How It Works

You (natural language) --> Claude --> MCP Server --> Fabric REST API --> Your Fabric Resources

Two transport modes:

• STDIO (default) - fabric_mcp_stdio.py - used by Claude Code and Claude Desktop
• HTTP - fabric_mcp.py --port 8081 - for custom integrations

Tools (77+)

Workspace & Resource Management (9 tools)

list_workspaces, create_workspace, set_workspace, list_lakehouses, create_lakehouse, set_lakehouse, list_warehouses, create_warehouse, set_warehouse

Delta Table Operations (10 tools)

list_tables, set_table, get_lakehouse_table_schema, get_all_lakehouse_schemas, table_preview, table_schema, describe_history, optimize_delta, vacuum_delta, load_data_from_url

SQL (4 tools)

sql_query, sql_explain, sql_export, get_sql_endpoint

• Always pass type ("lakehouse" or "warehouse")
• SQL endpoints are read-only - no DDL/DML
• New delta tables take 5-10 min to appear in SQL endpoint
• Requires ODBC Driver 18

Semantic Models & DAX (9 tools)

list_semantic_models, get_semantic_model, get_model_schema, list_measures, get_measure, create_measure, update_measure, delete_measure, analyze_dax_query

Reports (4 tools)

list_reports, get_report, report_export, report_params_list

Power BI (2 tools)

semantic_model_refresh, dax_query

Notebooks (16 tools)

list_notebooks, create_notebook, get_notebook_content, update_notebook_cell, create_pyspark_notebook, create_fabric_notebook, generate_pyspark_code, generate_fabric_code, validate_pyspark_code, validate_fabric_code, analyze_notebook_performance, run_notebook_job, get_run_status, cancel_notebook_job, install_requirements, install_wheel, cluster_info

Pipelines & Scheduling (8 tools)

pipeline_run, pipeline_status, pipeline_logs, create_data_pipeline, get_pipeline_definition, dataflow_refresh, schedule_list, schedule_set

OneLake File Operations (7 tools)

onelake_ls, onelake_read, onelake_write, onelake_rm, onelake_create_shortcut, onelake_list_shortcuts, onelake_delete_shortcut

Items & Permissions (4 tools)

resolve_item, list_items, get_permissions, set_permissions

Microsoft Graph (8 tools)

graph_user, graph_mail, graph_teams_message, graph_teams_message_alias, graph_drive, save_teams_channel_alias, list_teams_channel_aliases, delete_teams_channel_alias

Session (1 tool)

clear_context

Example Usage

"List all my Fabric workspaces"
"Set workspace to Analytics-Prod"
"Show me all tables in the sales lakehouse"
"What are the top 10 customers by revenue?"
"Create a DAX measure for total sales"
"Generate a PySpark ETL notebook"
"Run the daily pipeline and check status"

Manual Setup (without setup.py)

Claude Code (VS Code)

Create .mcp.json in the project root: ``json { "mcpServers": { "ms-fabric-core-tools-mcp": { "command": "uv", "args": ["--directory", "/full/path/to/ms-core-mcp", "run", "fabric_mcp_stdio.py"] } } } ``

Add to ~/.claude/settings.json: ``json { "enableAllProjectMcpServers": true } ``

Claude Desktop

Add to your claude_desktop_config.json:

• Windows: %APPDATA%\Claude\claude_desktop_config.json
• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "ms-fabric-core-tools-mcp": {
      "command": "/full/path/to/uv",
      "args": ["--directory", "/full/path/to/ms-core-mcp", "run", "fabric_mcp_stdio.py"]
    }
  }
}

Note: Claude Desktop needs the full path to uv since it may not inherit your shell PATH.

Project Structure

ms-core-mcp/
  fabric_mcp.py          # HTTP server entry point
  fabric_mcp_stdio.py    # STDIO entry point (Claude Code / Desktop)
  setup.py               # Automated setup script
  pyproject.toml          # Dependencies
  CLAUDE.md              # AI assistant instructions
  tools/                 # MCP tool definitions
  helpers/
    clients/             # Fabric API, SQL, OneLake clients
    formatters/          # Output formatters
    utils/               # Auth, context, validators
  tests/                 # Tests
  docs/                  # Documentation

Known Limitations

• SQL endpoints are read-only - use PySpark notebooks for writes
• SQL sync delay - new tables take 5-10 min to appear
• API rate limits - 50 requests/min/user (Fabric), 120 queries/min/user (Power BI)
• Notebook lakehouse attachment - must be done manually in Fabric UI
• ODBC Driver 18 required for sql_query, table_preview, sql_explain tools
• Report creation not supported - can list, export, view params only

Troubleshooting

"Command not found: uv" - Install uv, then restart terminal

"Not authenticated" - Run az login, then verify with: ``bash az account get-access-token --resource https://api.fabric.microsoft.com/ ``

"Database not found" on SQL queries - SQL endpoint may not be provisioned yet. Wait a few minutes or check the lakehouse has tables in the Fabric portal.

MCP tools not appearing in Claude Desktop - Use full path to uv in config. Check Task Manager to fully close and restart Claude Desktop.

Only some tools showing - Check for import errors by running: ``bash uv run python -c "from tools import *; from helpers.utils.context import mcp; print(f'Tools: {len(mcp._tool_manager._tools)}')" ``

Testing

uv run pytest tests/

License

MIT