Starburst MCP

Starburst's MCP server runs on the cluster coordinator and gives AI clients governed access to federated data. The query tool executes read only SQL and returns results as JSON with column metadata alongside the rows. Read only is enforced rather than advised: statements such as INSERT, UPDATE, DELETE, MERGE, TRUNCATE, GRANT and REVOKE are rejected, as is DDL, so an agent cannot mutate a warehouse or change permissions. Result size and query execution time are both capped by configurable limits, which stops one open ended question from consuming cluster capacity. Discovery tools address the harder problem of an agent not knowing what data exists. Data product search runs keyword search across data product names, summaries, descriptions, views and materialized views, and a details tool returns the full definition of one data product including its views and their schemas. Parameterized query tools let a team publish vetted query templates: one tool lists the available templates and another executes a template with validated parameters, which is a practical way to give agents useful access without handing them free form SQL. The Starburst AI agent is also reachable through the same interface for stateful interactions within a data product. The server is enabled through coordinator configuration and served at a path on the coordinator host, with OAuth 2.0, delegated OAuth or standard HTTP authentication headers, and it requires a valid license. Equivalent functionality is documented for Starburst Galaxy. It targets data engineers and analytics teams exposing governed data to AI applications.