Improves the reliability of AI-driven business analysis by enforcing data definitions and permissions across connected workplace tools.
What is Databricks Genie One MCP?
Databricks launched Genie One MCP on September 22, 2026, and the company describes it as governed context for any MCP-compatible AI agent. It connects assistants such as ChatGPT, Claude, Microsoft Copilot, and coding agents to Genie One, Databricks’ natural-language analytics service, so external agents query internal data without inventing what the business means.
The launch matters because general-purpose assistants already touch enterprise data, but without business context they misread metrics, trust the wrong sources, and ignore access controls. Genie One went generally available with this release, and setup starts in a workspace with the preview toggle plus a client connection.
Independent coverage the same day confirmed the availability claim: the MCP server is open to all Databricks users, with connections managed through Unity Gateway for centralized access controls, fine-grained policies, and audit logging.
Your AI agents can now carry your approved definitions with them, instead of inventing their own.
How does Genie One MCP solve the context problem?
Ungrounded assistants inherit the fragmentation of the systems they connect to. Relevant data sits in different systems, refreshes on different schedules, and changes over time, so a consolidated view can be stale before anyone reads it.
Definitions diverge too. Net sales and active promotion mean different things to sales, finance, and marketing, and the working definitions often live in spreadsheets or institutional knowledge rather than structured code. When numbers conflict, it is unclear which source wins.
Genie Ontology attacks that gap by defining approved metrics, data relationships, source authority, and permissions once, then serving the same context to every approved agent. The Model Context Protocol is the open standard that carries that context into the tools teams already use.
Define the business once, and every agent inherits the same definition instead of 3 competing ones.
How is Genie One MCP different from connecting an agent straight to your data?
Direct connections provide access without shared context, and the costs show up in accuracy, latency, and governance. Without a governance layer, agents inspect schemas and documentation again and again, which burns tokens and slows every answer.
Each client, model, or session can land on a different interpretation of the same definition, so the same question returns different answers on different days. Genie One reconciles conflicts based on authority and certification inside the ontology, and the managed MCP documentation shows how each connection maps to Unity Catalog governance rather than raw table access.
The consumption model follows the feature you call: Genie Agents run on serverless SQL compute pricing, Unity Catalog functions on serverless general compute, and SQL servers on Databricks SQL pricing. There is no flat fee for the MCP layer itself, so the cost scales with how hard your agents query.
Direct access answers fast and wrong, governed access answers right, and only one of those survives a board meeting.
Who is Genie One MCP actually for?
The feature targets data teams and the executives who consume their numbers, in organizations already running Databricks with Unity Catalog. If your metrics live in one warehouse with certified definitions, the MCP layer extends that trust to external agents.
Smaller teams without a data platform face a different reality, and pulling 130-plus tools into one governed dashboard delivers a version of the same discipline without enterprise infrastructure.
Enterprise data platforms need this first, and small teams need the same discipline at smaller scale.
The Monday report shows 3 different net sales figures, one from the sales dashboard, one from finance, and one from the AI assistant your ops lead swears by.
Each number is correct inside its own system and none of them agree, which is the exact gap Genie Ontology closes before an agent writes a query.
Databricks caps each Genie SQL run at 90 seconds, but no timeout fixes a definition that was wrong before the query started.
Is Genie One MCP production ready?
Generally available, with real constraints. The recommended setup uses on-behalf-of OAuth, where the external assistant passes the end user’s token and Genie evaluates Unity Catalog privileges, row filters, and column masks in that user’s context, per the client connection documentation.
Machine-to-machine authentication with a service principal is available for external integrations, but it represents every caller as one identity and removes per-user permission enforcement. Account for the 90-second SQL execution timeout and the workspace Genie QPM limit when sizing a rollout, and note that chat events land in audit logs, SQL execution in Query History, and consumption in billing system tables.
The same documentation set covers the client side, and Claude, Claude Code, Cursor, and Replit all connect through OAuth apps scoped to the ai-gateway rather than broad all-apis tokens. External MCP connections operate as Unity Catalog objects governed through standard grants, so your existing access reviews apply to the agents too.
Production-ready for governed deployments, and a liability when every caller shares one identity.
Source: Databricks Blog
