LLMs are vastly more capable, but they are not databases. An LLM is a probabilistic engine, not guaranteed to return the same answer twice, even to an identical question. Yet many teams let LLMs generate raw text-to-SQL at runtime, a risky anti-pattern that invites hallucinated columns, performance blindspots, and unpredictable business logic.

This talk introduces Google's open-source MCP Toolbox for Databases, built natively on the Model Context Protocol. Instead of letting an LLM guess your schema or improvise syntax, Toolbox exposes structured, pre-defined operations across PostgreSQL, BigQuery, Oracle, and many other database engines, so the model chooses what to call, never how to query.

Live demo: wiring Toolbox into an agent so it retrieves enterprise data safely and predictably, no raw query generation exposed to the model.

You'll walk away with a concrete blueprint for making your database a secure, predictable citizen in an LLM-powered pipeline.