MCP Toolbox for Databases is an open-source database MCP server from Google, designed to connect AI agents, IDEs, and applications to enterprise databases. The project was initially named "Gen AI Toolbox for Databases" (repository genai-toolbox), and because early development predated the MCP protocol, it was later renamed mcp-toolbox to reflect its compatibility with MCP.
[Purpose] This tool has a dual purpose: first, as an out-of-the-box MCP server (build-time), it can immediately connect MCP clients such as Gemini CLI, Google Antigravity, Claude Code, and Codex to databases, using prebuilt generic tools for data querying, schema exploration, and code generation without writing boilerplate code. Second, as a custom tool framework (runtime), it provides the ability to build specialized, highly secure AI tools, supporting the definition of structured queries, semantic search, and NL2SQL, suitable for production-grade agents.
[Main Capabilities]
- Out-of-the-box database access: prebuilt generic tools (such as list_tables, execute_sql) allow direct data exploration from an IDE or CLI.
- Custom tool framework: define sources, tools, toolsets, and prompts through the tools.yaml configuration file to achieve restricted access, structured queries, and semantic search.
- Simplified development: can be integrated into Agent Development Kit (ADK), LangChain, LlamaIndex, or custom agents within 10 lines of code.
- Performance and observability: built-in connection pooling, IAM integrated authentication, and end-to-end observability (OpenTelemetry), providing out-of-the-box metrics and tracing.
- Enhanced security: integrated authentication for more secure data access.
[Supported Databases] Google Cloud: AlloyDB, BigQuery, Cloud SQL (PostgreSQL, MySQL, SQL Server), Spanner, Firestore, Knowledge Catalog (formerly Dataplex). Other databases: PostgreSQL, MySQL, MariaDB,SQL Server, Oracle, MongoDB, Redis, Elasticsearch, CockroachDB, ClickHouse, Couchbase, Neo4j, Snowflake, Trino, etc.
[Usage]
- Installation and running: supports binaries, Docker containers, Homebrew, and source compilation, and can also be run directly via npx (requires Node.js). After installation, use the toolbox command to start the server, use --config to specify the configuration file, --prebuilt=<database> to load prebuilt tools, and --prebuilt=<database>/<toolset> to load a specific toolset.
- Connecting an MCP client: configure the MCP server in mcp.json or claude_desktop_config.json. For example: { "mcpServers": { "toolbox-postgres": { "command": "npx", "args": ["-y", "@toolbox-sdk/server", "--prebuilt=postgres", "--stdio"] } } } After that, tools can be loaded into a compatible MCP client or application; when connecting to a specific toolset, use the URL http://127.0.0.1:5000/mcp/{toolset_name}。
- Custom tools: define sources, tools, toolsets, prompts, resources, etc. through tools.yaml, and run toolbox to start the server. Dynamic reloading is supported (enabled by default and can be disabled with --disable-reload).
- SDK integration: provides client SDKs such as Python (toolbox-core, toolbox-langchain, toolbox-llamaindex), JavaScript/TypeScript (@toolbox-sdk/core, @toolbox-sdk/adk), and Go (mcp-toolbox-sdk-go), making it easy to load toolsets in applications.
[Applicable Scenarios] Suitable for development scenarios that require connecting AI agents or IDEs to databases, including data exploration, schema management, code generation, building production-grade database AI tools, and natural language querying of databases.
[Limitations] Specific performance metrics, charging models, pricing, and platform compatibility details are not yet provided. The specific implementation details of some advanced features (such as semantic search and NL2SQL) require reference to the official documentation.
