Databricks Launches Native Genie Plugins for Four AI Agents

By Grace Kim

Databricks launched native Genie plugins on November 11, 2026, for ChatGPT, Claude, Microsoft Copilot Cowork and Grok Bot. The release gives all four agents governed access to business context held in Databricks, while other compatible agents can connect through the Genie One MCP server.

Databricks had already connected Genie to outside assistants before the launch. Its September 22, 2026 announcement described governed context for assistants and coding agents through Model Context Protocol, or MCP, a protocol for connecting AI agents to external tools. The November plugins add native interfaces inside the four named products while using the same context layer.

How does Genie give agents governed business context?

Databricks says Genie Ontology combines deliberately defined metrics and business relationships with context inferred from governed tables, queries, dashboards, notebooks and other enterprise assets. It ranks that material by authority and relevance, applies permissions, and supplies the definitions, relationships and source context Genie uses to interpret a question.

An organization can maintain an approved metric definition once and apply it across connected agents. When a question arrives, Genie resolves business terms, finds data the user may access, writes and runs SQL, and returns an answer with citations to Databricks sources.

Genie can send specialized work to domain-specific Genie Agents, including marketing for pipeline attribution and finance for operational costs. It can also draw insights from unstructured enterprise data.

An agent can hand Genie an entire business question or request only the context it is missing before continuing with its own tools, such as editing files, drafting plans or posting to Slack.

Clients supporting the MCP Apps extension can display Genie’s progress, visualizations, results and ontology citations in an interactive panel. Other clients continue to receive text results.

The practical result is a shared definition layer that can travel across agent interfaces without surrendering access controls.

What can each agent do with the connection?

Databricks’ examples separate analysis, document production and campaign execution across the four native interfaces.

• ChatGPT Dot: The always-on agent, which has its own computer, can monitor feature adoption against targets using a certified active-user definition, pull related support themes and post a summary to Slack.
• Claude: It can investigate a regional revenue shortfall with certified finance metrics, pricing memos, sales-call notes and competitor reports, then produce an analysis with citations behind each claim.
• Microsoft Copilot Cowork: It can use account and usage trends from Azure Databricks to populate Word briefs, Excel charts and client presentations, then wait for approval before sending anything.
• Grok Bot: As a growth teammate, it can review campaigns using governed definitions for campaign spend, qualified opportunities and pipeline, recommend budget shifts with evidence, and prepare retargeting or lifecycle campaigns after a growth lead approves them.

Across all four workflows, Genie supplies the governed business numbers while the agent handles documents, communications or campaign execution.

How are permissions and authentication enforced?

Databricks says Unity Catalog permissions are enforced on every request, so results remain limited to the data each user can access.

The Genie One MCP documentation recommends on-behalf-of user OAuth authentication, which carries the end user’s identity into the request. It also supports service-principal authentication for automated workloads.

Setup starts by enabling the plugin in the agent workspace. The ChatGPT and Claude plugins use default OAuth apps, allowing users to sign in with their Databricks accounts without an administrator creating an OAuth application.

Genie responses are asynchronous because the system searches data and executes SQL.

Standard response tools truncate query results to protect the model’s context window. A separate tool retrieves the full schema and rows, although very large results can still be truncated.

The practical takeaway is simple: every answer passes through Unity Catalog permissions before Genie returns data.

What must customers change before the deadline?

The older beta MCP endpoint reaches end of life on November 30, 2026. Databricks says requests to it will fail after that date, requiring affected workloads to move to the generally available Genie One MCP endpoint.

Answer quality depends on the data estate behind the ontology. Databricks’ September 1, 2026 implementation guide says: “Genie Ontology works on day one.” The company adds that higher accuracy depends on clean data, consistent entity records, useful metadata, governed metric definitions, certified assets and continuing evaluation.

The guide recommends starting with one domain, then strengthening its definitions, sources, governance and evaluation before expanding.

The launch gives Databricks a consistent governance layer across agents with different native strengths. As those agents move from cited analysis to approved actions, the shared ontology can determine which numbers they use and which underlying enterprise data each person can see.