MCP Toolbox for Databases

Overview

MCP Toolbox for Databases (repository: googleapis/mcp-toolbox) is Google's open-source Model Context Protocol (MCP) server designed specifically for databases. It allows AI agents, IDEs, and applications to interact directly and securely with enterprise data sources using natural language and structured tools—no custom integration code required for most scenarios.

Formerly known as Gen AI Toolbox, it was renamed to better reflect its deep integration with the MCP ecosystem. Tools are defined declaratively in a tools.yaml file, and the server handles connection management, authentication, query execution, and security.

Key Features

• Broad Database Support: Native integration with Google Cloud (BigQuery, Cloud SQL for PostgreSQL/MySQL/SQL Server, AlloyDB, Spanner, Firestore) and many third-party sources including PostgreSQL, MySQL, Oracle, MongoDB, Redis, Elasticsearch, Snowflake, Neo4j, Trino, and more.
• Prebuilt Generic Tools: Ready-to-use tools like list_tables, execute_sql, schema exploration, and data querying that work immediately.
• Custom Tools Framework: Define structured queries, semantic search, NL2SQL, and domain-specific tools via configuration.
• Enterprise Security: IAM-based authentication, restricted access, parameterized queries, and least-privilege execution.
• Observability: Built-in OpenTelemetry support for metrics and tracing.
• Dynamic Reloading & Toolsets: Update tools without restarting; group tools into modular toolsets.
• Interactive UI: Built-in web UI for testing tools (enabled with --ui flag).
• SDK Ecosystem: Official client SDKs for Python, JavaScript/TypeScript, Java, and Go to integrate Toolbox tools into agent frameworks like LangChain, LlamaIndex, and Genkit.
• Agent Skills Generation: Convert toolsets into portable Agent Skills.

How It Works

1. Configure data sources and tools in tools.yaml (sources, tools, toolsets, prompts).
2. Start the MCP Toolbox server (binary, Docker, or cloud deployment). It exposes an MCP endpoint (default: http://127.0.0.1:5000/mcp).
3. MCP clients (Gemini CLI, Claude Code, Cursor, etc.) discover and call the tools.
4. The server executes operations against the databases with proper auth and pooling, returning structured results to the agent.

Supports both HTTP-based MCP connections and prebuilt tool configurations for quick setup.

Use Cases

• Natural Language Data Querying: AI assistants query and analyze live data using plain English.
• Schema-Aware Development: IDEs and agents explore schemas and generate accurate database code.
• Secure Enterprise AI Agents: Build production agents that interact with sensitive data under controlled access.
• Automated Code & Query Generation: Generate SQL, migrations, or application logic that matches real schemas.
• Data Exploration & Analysis: Combine with LLMs for insights, reporting, or debugging data issues.
• Multi-Database Workflows: Unify access across heterogeneous data sources in one MCP server.

Getting Started

Installation Options

• Binary: Download latest release from GitHub or Google Cloud Storage.
• Homebrew: brew install mcp-toolbox
• Docker: Pull from us-central1-docker.pkg.dev/database-toolbox/toolbox/toolbox
• Go: go install github.com/googleapis/mcp-toolbox@latest
• Prebuilt via npx: Quick setup for specific databases.

Run with:

./toolbox --config tools.yaml

Add to your MCP client config (example for HTTP mode or prebuilt):

{
  "mcpServers": {
    "toolbox": {
      "type": "http",
      "url": "http://127.0.0.1:5000/mcp"
    }
  }
}

Full documentation: https://mcp-toolbox.dev/

Benefits

MCP Toolbox significantly lowers the barrier to building data-aware AI agents while meeting enterprise security, scalability, and observability requirements. It is production-ready (v1.0.0 released April 2026), actively maintained by Google, and serves as a reference implementation for secure MCP database integrations.

With support for dozens of data sources and a rich SDK ecosystem, it enables developers to create powerful, reliable agentic applications that operate on real enterprise data.