// the find
googleapis/mcp-toolbox
MCP Toolbox for Databases is an open source MCP server for databases.
MCP Toolbox is a Google-backed MCP server that acts as a middleware layer between AI agents and databases. It lets you define named SQL tools in a YAML file and expose them over HTTP to any MCP-compatible client, with prebuilt generic tools for instant schema exploration. Aimed at teams building production AI agents that need controlled, auditable database access rather than raw SQL from an LLM.
The YAML-defined tool approach is the right call — you get parameterized queries with explicit schemas instead of letting the LLM construct raw SQL, which matters for both safety and predictability. Connection pooling and OpenTelemetry are built in, so you're not bolting those on yourself later. SDK coverage is genuinely broad (Python, JS, Go, Java) with framework-specific adapters for LangChain, LlamaIndex, Genkit, and ADK. The prebuilt tools mode via npx is a zero-config path that actually works for IDE-connected exploratory use.
The tools.yaml config format is its own DSL that you have to learn and maintain alongside your actual codebase — schema changes in Postgres mean updating both your migrations and your toolbox config, which will drift. The server is a separate process you have to run, deploy, and monitor; it adds an extra network hop and failure point compared to embedding database access directly in your agent. Authentication support exists but is Google-IAM-first; non-GCP deployments get generic OAuth which requires more integration work. Dynamic SQL via NL2SQL is buried and the docs don't make it clear where the guardrails actually are on arbitrary query generation.