Snowflake MCP

Snowflake Labs publishes an MCP server that exposes a Snowflake account to AI clients across three areas: Cortex AI services, object management and SQL orchestration. Cortex Search handles unstructured data for retrieval augmented workflows, Cortex Analyst answers questions over structured data through semantic models, and a Cortex Agent tool orchestrates across both. Semantic views defined in the account can also be discovered and queried directly. Beyond the AI services it does ordinary warehouse work. An agent can create, alter and drop the common objects, meaning databases, schemas, tables, views and warehouses, and it can execute generated SQL. The configuration file is the part that matters most. You declare which services are exposed, and a statement permissions section lists which categories of SQL statement are allowed, so destructive operations can be refused by configuration rather than left to how carefully a prompt was worded. That design makes it usable in places where handing a model raw warehouse credentials would never clear review. Analysts get natural language querying that resolves through governed semantic models instead of guessed column names, and platform teams keep a declarative boundary around what any connected client can do. Pair it with a role scoped to the data you intend to expose, since MCP level permissions layer on top of Snowflake's own grants rather than replacing them.