Cube MCP

Cube's MCP server puts a governed semantic layer between an AI client and the data warehouse, so questions about business metrics resolve against defined models rather than raw tables. The agent can list the cubes and views available to it, describe the measures, dimensions, and segments each one carries, and then run queries with filters, time dimensions, and a chosen granularity, receiving results already shaped by the modeling layer. That indirection is the entire point. A language model writing SQL straight against a warehouse invents joins and picks the wrong revenue column; a model querying through Cube gets the single approved definition of revenue, active user, or churn, with access policies and row level security applied on the way through. Answers stay consistent with the dashboards built on the same models, and the query the agent issued remains inspectable when a number looks wrong. It fits data teams that already maintain Cube models and want analysts or an internal chat assistant to self serve without direct warehouse access, and product developers embedding conversational analytics in an application. Because the server speaks the standard protocol, one set of models can back a coding assistant, a BI chat interface, and an automated agent workflow without building a separate integration for each of them.