Azure Databricks exposes Genie One’s governed text-to-SQL through MCP

Azure Databricks has introduced a beta managed Model Context Protocol server for Genie One, giving external MCP clients and agents a direct route into its conversational analytics system. An agent can send a natural-language question to the workspace endpoint; Genie searches permitted enterprise data, writes SQL and returns an answer grounded in Genie Ontology, with links back to the cited Databricks sources. Microsoft’s documentation says Unity Catalog permissions are enforced on every request.

What the endpoint actually exposes

The server is not a generic database connection that leaves SQL generation to the calling model. Its main genie_ask tool starts a Genie response, while separate tools poll the asynchronous job, retrieve the full query result or cancel a turn. Follow-up questions can reuse the same conversation ID. This keeps business terms, metric definitions and table relationships inside Genie’s governed ontology rather than asking an outside agent to infer them from raw schemas. The documented endpoint and tool contract also let callers specify the SQL warehouse used to execute a request.

For clients that implement MCP Apps, Azure Databricks can return an interactive embedded view showing Genie’s progress, visualizations and final answer. Other clients receive text results and source links. Microsoft’s example shows the view inside Claude Desktop, and the connection guidance also names Cursor as a supported external client. The same guide says the server uses workspace authentication and recommends on-behalf-of-user OAuth with the genie scope.

Governance travels with the query

That authentication model is the important part for enterprise text-to-SQL deployments. Unity Catalog permissions remain enforced on every request; with the recommended on-behalf-of-user OAuth flow, results can be scoped to the user’s existing access instead of relying on a shared, broadly privileged warehouse identity. The server also honors the workspace’s Genie One configuration, so teams tune analytics behavior in Databricks while making the resulting interface available to outside assistants. Microsoft’s considerations section notes that normal responses truncate query results to protect the model context window; agents can request the complete schema and rows through a separate tool, still subject to a size limit.

The release is deliberately broad in scope: it exposes Genie One across a workspace. Databricks points teams that need a narrower, curated analytics domain to the separate Genie Agent MCP server. The beta documentation requires an administrator to enable the Managed MCP Servers workspace preview before the endpoint becomes available.

The move arrives alongside a wider MCP expansion in the product. Azure Databricks’ September 10 release notes say Databricks-provided connectors for services including Google Drive, Microsoft 365, Slack and GitHub are now generally available for Genie One and Genie Code, governed through Unity Catalog and Unity Gateway. Together, the releases make MCP bidirectional: Genie can call external tools, while external agents can now call Genie for governed natural-language analytics.