Genie ↔ Fabric/Power BI bridge (MCP server)

Custom MCP server, deployed as a Databricks App, that lets an MCP client (Claude, Cursor, etc.) ask questions against Databricks Genie, and check that the Fabric/Power BI side of the pipeline is healthy.

Read this alongside the architecture diagram shared in chat before deploying.

What this does — and doesn't — automate

The data model moves from Databricks to Power BI via Fabric mirroring, not via MCP:

1. Your data model lives in Databricks Unity Catalog.

2. Fabric's Mirrored Azure Databricks Catalog creates zero-ETL OneLake shortcuts to it and keeps them in sync automatically.

3. Power BI reads that data live via a Direct Lake semantic model.

4. A person uses Power BI Copilot, in the Power BI UI, to generate the actual dashboard/report from a natural-language prompt.

This MCP server sits alongside that pipeline. It does not create the mirroring (set up once, manually, in the Fabric portal — see below), and it cannot trigger Copilot's report generation (no public API exists for that). What it does give an MCP client:

• genie_ask / genie_get_result / genie_list_spaces — ask Genie questions and read back answers, generated SQL, and result rows.

• fabric_mirror_status — confirm the mirrored catalog is synced.

• powerbi_list_datasets / powerbi_list_reports / powerbi_refresh_dataset — discovery and refresh operations against Power BI.

• genie_to_copilot_handoff — turns a Genie answer into a ready-to-paste Copilot prompt, so the human's last step is "paste and go" instead of starting from scratch.

Related MCP server: Microsoft Fabric MCP Server

Prerequisites (set these up before deploying the server)

1. Genie space in Databricks over your Unity Catalog data model. Note its space ID.

2. Fabric mirroring: in the Fabric portal, create a Mirrored Azure Databricks Catalog item pointing at your Unity Catalog. Note the Fabric workspace ID and the mirrored item ID. https://learn.microsoft.com/en-us/fabric/mirroring/azure-databricks-tutorial

3. Power BI semantic model in Direct Lake mode built on top of the mirrored OneLake tables.

4. Entra ID app registration (service principal) for this server to call the Power BI REST API and Fabric REST API:

   • Grant it a Power BI service principal profile / workspace access (add it as a member of the target Power BI/Fabric workspace).

   • Note its tenant ID, client ID, and client secret.

5. Databricks CLI / workspace access to deploy a Databricks App.

Deploying as a Databricks App

1. Register the values from step 4 as Databricks secrets (scope name is up to you), matching the valueFrom keys in app.yaml: genie-space-id, powerbi-tenant-id, powerbi-client-id, powerbi-client-secret, powerbi-group-id, fabric-workspace-id, fabric-mirrored-item-id.

2. Create the Databricks App and upload this folder's source.

3. Deploy. The MCP endpoint will be available at https://<app-url>/mcp.

4. Point your MCP client at that URL (OAuth is handled by Databricks Apps for inbound access).

Full steps: https://docs.databricks.com/aws/en/generative-ai/mcp/custom-mcp

Local development

pip install -r requirements.txt
cp .env.example .env   # fill in real values
python app.py

Before you deploy this for real — verify these against current docs

I wrote this scaffold from public documentation, not against a live workspace, so a few pieces need a quick check before you trust them:

• src/genie_tools.py: the databricks-sdk Genie method names (start_conversation_and_wait, create_message_and_wait, get_message_query_result, list_spaces) — confirm against the version of databricks-sdk you pin.

• src/fabric_tools.py: the Fabric REST API path for item status — Fabric's API surface is newer and evolves faster than Power BI's.

• app.yaml: the secret-binding syntax (env / valueFrom) — confirm against the current Databricks Apps schema.

• app.py: the mcp.run(...) call signature for the installed mcp SDK version (host/port handling has changed across releases).

None of these are exotic, but I can't execute this against your workspace, so treat this as a solid first draft to test, not a verified-working build.