MCP Toolbox for Databases Documentation
by Google
Give AI agents safe, pooled, observable database tools over MCP, defined in a single tools.yaml.
Overview
MCP Toolbox for Databases (formerly GenAI Toolbox, repo googleapis/genai-toolbox) is an open-source server from Google that sits between an agent's orchestration framework and its databases as a control plane for defining, sharing and invoking tools. The documentation at mcp-toolbox.dev is organised into Getting Started (introduction, an MCP quickstart, and local Python, JavaScript and Go quickstarts), Configuration (sources, tools, toolsets and groups; authentication with Google Sign-In or generic OIDC; security with Model Armor and read-only tools; prebuilt configs; embedding models; prompts; resources; MCP Apps; Toolbox UI; agent skills; pre- and post-processing), Connect to Toolbox (MCP clients, Python/JavaScript/Go/Java SDKs, Gemini CLI extensions and IDE setups), Deploy (Cloud Run, Docker Compose, Kubernetes), Monitoring (OpenTelemetry telemetry and SQL Commenter), Integrations for 40+ databases, samples and a reference section. Tools are declared in a tools.yaml file and served with ./toolbox --config tools.yaml, and an optional --ui flag opens a testing interface. The Python quickstart builds a hotel search-and-booking agent on PostgreSQL 16+ and shows the same tools in Google ADK, LangChain/LangGraph, LlamaIndex and the core SDK with Gemini, Claude or Vertex AI. The security page covers prompt-injection screening with Model Armor and read-only enforcement through engine-level protocol locks, tool suppression and MCP annotations. The project releases weekly: v1.11.0 (10 September 2026) added MCP Apps and Resources, and v1.13.1 shipped on 25 September 2026.
At a Glance
- Topic
- MCP
- Level
- Intermediate
- Format
- Documentation
- Cost
- Free
- Duration
- ~1 hour for a local quickstart; roughly half a day self-paced to cover configuration, security and deployment (estimate)
- Provider
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓How to declare database sources, tools, toolsets and groups in a single tools.yaml configuration file
- ✓How to run the Toolbox server locally and test tools with the built-in Toolbox UI
- ✓How to load toolsets into Google ADK, LangGraph, LlamaIndex or the core SDK from Python
- ✓How to connect Toolbox to MCP clients such as Gemini CLI and Claude Code for database access
- ✓How to secure agent database access with OIDC authentication, read-only modes and Model Armor screening
- ✓How to deploy Toolbox to Cloud Run, Docker Compose or Kubernetes for production agents
- ✓How to export OpenTelemetry traces and metrics and tag queries with SQL Commenter
- ✓How to use prebuilt configs and generic tools like list_tables and execute_sql across 40+ databases
Highlights
- •Very active project: about 16.5k GitHub stars and weekly releases, v1.10 through v1.13.1 between 27 Aug and 25 Sep 2026 (checked 2026-09-26)
- •Handles the production concerns most DIY database tools skip: connection pooling, IAM/OIDC auth, read-only enforcement and OpenTelemetry
- •Wide coverage, from AlloyDB, BigQuery and Spanner to PostgreSQL, MySQL, MongoDB, Redis, Neo4j, Snowflake, ClickHouse and Trino
- •Framework-neutral: SDKs for Python, JavaScript/TypeScript, Go and Java, plus any MCP client
- •The quickstart runs the same hotel-booking tools in four agent frameworks, which makes it easy to compare them
Who It’s For
Best For
- ✓Engineers building data agents or text-to-SQL assistants on existing databases
- ✓Platform teams that want one governed MCP server for database tools instead of per-agent code
- ✓Developers wiring Gemini CLI, Claude Code or IDE agents to their databases
- ✓Google Cloud users on AlloyDB, BigQuery, Cloud SQL or Spanner
Prerequisites
- •Basic SQL and access to a database such as PostgreSQL 16+ (the quickstart's default)
- •Python 3.10+, Node.js or Go, depending on which SDK quickstart you follow
- •Familiarity with agents, tool calling and MCP concepts, plus an LLM API key (Gemini, Claude or Vertex AI)
FAQ
What is MCP Toolbox for Databases Documentation?
The MCP Toolbox for Databases documentation is Google's guide to an open-source MCP server that connects AI agents to PostgreSQL, BigQuery, MongoDB, Snowflake and dozens of other databases. It is for engineers building data agents. Afterwards you can define, secure, deploy and observe production database tools for any MCP client or agent framework.
Is MCP Toolbox for Databases Documentation free?
MCP Toolbox for Databases Documentation is free to access.
What level is MCP Toolbox for Databases Documentation for?
MCP Toolbox for Databases Documentation is aimed at a intermediate audience. Recommended background: Basic SQL and access to a database such as PostgreSQL 16+ (the quickstart's default), Python 3.10+, Node.js or Go, depending on which SDK quickstart you follow, Familiarity with agents, tool calling and MCP concepts, plus an LLM API key (Gemini, Claude or Vertex AI).
How long does MCP Toolbox for Databases Documentation take?
Expect roughly ~1 hour for a local quickstart; roughly half a day self-paced to cover configuration, security and deployment (estimate). Most learners work through it at their own pace.
What will I learn from MCP Toolbox for Databases Documentation?
You'll learn: How to declare database sources, tools, toolsets and groups in a single tools.yaml configuration file; How to run the Toolbox server locally and test tools with the built-in Toolbox UI; How to load toolsets into Google ADK, LangGraph, LlamaIndex or the core SDK from Python; How to connect Toolbox to MCP clients such as Gemini CLI and Claude Code for database access; How to secure agent database access with OIDC authentication, read-only modes and Model Armor screening; How to deploy Toolbox to Cloud Run, Docker Compose or Kubernetes for production agents; How to export OpenTelemetry traces and metrics and tag queries with SQL Commenter; How to use prebuilt configs and generic tools like list_tables and execute_sql across 40+ databases.
Topics
Sources
This page was written from 6 sources, 2 on domains other than mcp-toolbox.dev.