Moving from Raw SQL to Governed Data Models
Building enterprise AI agents that interact directly with raw databases often leads to fragile, unmaintainable SQL. The core challenge is that LLMs lack context regarding business logic, definitions, and access controls. By bridging agents to Looker’s governed semantic layer, developers ensure that AI-driven insights are grounded in a single, validated source of truth. This architecture separates business logic (maintained in Looker) from execution (handled by the agent).
The Role of MCP and the Agent Development Kit
The Model Context Protocol (MCP) acts as a universal language, allowing agents to discover and invoke capabilities within Looker without needing to understand complex API calls.
- MCP Toolbox: An open-source, production-ready resource that exposes data sources as standardized tools. It includes a pre-built Looker toolset, allowing developers to perform tasks like auditing LookML project files or executing governed queries without writing custom integration code for every project.
- Agent Development Kit (ADK): This layer orchestrates the LLM and the agent. It allows the agent to discover tools via MCP and execute technical tasks. Because the MCP bridge stays in sync with LookML, any changes made to the underlying business model are automatically reflected in the agent's context.
Modular Architecture for Scalability
This approach provides a flexible, modular framework for building AI applications:
- Looker Layer: Centralizes business logic, dimensions, measures, and access filters.
- Transport Layer: Powered by MCP and the AI toolbox, it acts as a translator, exposing internal functions (metadata retrieval, query execution) as discoverable tools defined in a
tools.yamlconfiguration. - Application Layer: Utilizes the ADK to orchestrate the LLM, allowing developers to choose their preferred model while maintaining strict governance over the data the agent can access.