mcp-fabric — Microsoft Fabric MCP server

A FastMCP server for developing Fabric notebooks and Dataflow Gen2 through the Fabric REST API (https://api.fabric.microsoft.com/v1). Tools take a workspace (display name or GUID) and an item name/GUID; call list_workspaces first.

Read tools always work; create/update/delete/run tools require "writable": true in config.json.

Tools

Discovery / read

• list_workspaces
• list_items — filter by item_type (Notebook, Dataflow, Lakehouse, …)
• list_folders — workspace folders; pass a displayName/id as folder when creating items
• get_item — item metadata
• get_item_definition — decoded definition parts of any item (generic/advanced)

Notebooks

• create_notebook — from source (code string) or ipynb (full notebook JSON)
• get_notebook — returns the source extracted from the notebook's ipynb
• update_notebook — replace content
• run_notebook — run on demand with optional parameters → returns a jobInstanceId

Dataflow Gen2

• create_dataflow — from a Power Query M mashup_document
• get_dataflow — decoded parts (mashup.pq = the M query, queryMetadata.json)
• update_dataflow — replace the M document
• refresh_dataflow / publish_dataflow — on-demand jobs

Schedules

• get_item_schedules — schedules on an item: id, enabled, and recurrence

(Cron/Daily/Weekly, start/end, timezone). An empty list means the item only runs on demand or via a parent pipeline.

• set_schedule_enabled — flip one schedule's enabled flag by id
• disable_item_schedules — flip every schedule on an item (enable=True

to re-enable); schedules already in the target state are skipped

job_type selects which schedules you're looking at: Pipeline for data pipelines (default), RunNotebook for notebooks, Refresh for dataflows. Fabric's Update Schedule API replaces the whole schedule, so both write tools read the existing recurrence and echo it back — only the enabled flag changes.

Lakehouse / Warehouse data (read-only SQL) — fabric_sql.py, added 2026-10-08

• list_sql_tables — tables/views on a Lakehouse SQL endpoint, Warehouse or Mirrored database
• describe_sql_table — columns + types + row count of schema.table
• query_sql — one SELECT / WITH … SELECT, max_rows 200 by default (cap 5000), truncated flag
• refresh_sql_endpoint — re-sync a lakehouse endpoint's metadata (new / path-written tables, new shortcuts)
• list_onelake — list a OneLake folder (Tables, Tables/<schema>, Files/...); sees shortcuts and new tables before the SQL endpoint does

Read-only three ways: a statement check (single statement, must start SELECT/WITH, no DML/DDL/INTO/EXEC/SET/DECLARE), an ApplicationIntent=ReadOnly connection, and a rollback after every query. These tools work even when "writable": false. They run as this server's identity (login_hint in config.json) via a https://database.windows.net token and pyodbc + ODBC Driver 18. Lakehouse object names are case-sensitive (jde_prod.lot_sales). If a name is both a Lakehouse and something else, pass item_type (Lakehouse | Warehouse | MirroredDatabase).

Jobs / lifecycle

• get_job — status of a notebook run / dataflow refresh
• cancel_job
• move_item — move an item into a folder (or to the workspace root); children follow
• delete_item

Definition create/update are long-running operations; the server polls them to completion automatically. Notebook/dataflow runs return a jobInstanceId you monitor with get_job (they aren't polled to completion).

Known limitations (per Microsoft docs)

• Dataflow Gen2 run APIs: refresh/publish can be invoked, but Microsoft

currently notes the run may not complete successfully via API.

• Service-principal auth is not supported for dataflows (works for notebooks).

Auth

Set "auth" in config.json:

The identity needs an appropriate workspace role (Admin/Member/Contributor) to create and run items. Default token scope is https://api.fabric.microsoft.com/.default.

Two extra config.json keys tune broker sign-in when the machine has more than one work account signed in:

Setup

python -m venv .venv
.\.venv\Scripts\python.exe -m pip install -r requirements.txt
copy config.example.json config.json   # then edit if needed (broker auth works as-is)
.\.venv\Scripts\python.exe server.py    # smoke test (Ctrl+C to stop)

Register with an MCP client

See examples/mcp.json:

{
  "mcpServers": {
    "fabric": {
      "command": "C:\\path\\to\\mcp-fabric\\.venv\\Scripts\\python.exe",
      "args": ["C:\\path\\to\\mcp-fabric\\server.py"],
      "env": {}
    }
  }
}

Use with Claude Desktop

Claude Desktop reads its MCP servers from claude_desktop_config.json. Open it from Settings → Developer → Edit Config (this creates the file if it doesn't exist), or edit it directly:

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

Add this server under mcpServers, using absolute paths to the venv's Python and server.py:

{
  "mcpServers": {
    "fabric": {
      "command": "C:\\path\\to\\mcp-fabric\\.venv\\Scripts\\python.exe",
      "args": ["C:\\path\\to\\mcp-fabric\\server.py"],
      "env": {}
    }
  }
}

On macOS the paths are POSIX, e.g. "command": "/Users/you/mcp-fabric/.venv/bin/python". Save the file and fully quit and reopen Claude Desktop (use Quit from the tray/menu-bar icon — closing the window isn't enough). The server's tools then appear in the tools (🔌) menu of a new chat.

License

MIT — see LICENSE.