MCP Toolbox for Databases

MCP Toolbox for Databases is an open-source Model Context Protocol (MCP) server maintained by Google's googleapis organization, designed as a middleware control plane to connect AI agents, IDEs, and applications directly and securely to enterprise databases. The core service is built in Go (96.2%) under the Apache-2.0 license, currently at v1.1.0 following SemVer.

Dual-Mode Operation#

• Ready-to-Use Prebuilt Tools: Zero-code configuration via --prebuilt flag to instantly provide Gemini CLI, Claude Code, Codex and other MCP clients with common database exploration capabilities (e.g., list_tables, execute_sql).
• Custom Tool Framework: Declarative tools.yaml configuration defining four resource types — Source (data source), Tool (executable action), Toolset (tool grouping), and Prompt (prompt template) — to build production-grade tools with restricted access, structured queries, and semantic search.

Supported Databases#

Google Cloud Suite: AlloyDB, BigQuery, Cloud SQL (PostgreSQL/MySQL/SQL Server), Spanner, Firestore, Knowledge Catalog.

Third-Party Databases: PostgreSQL, MySQL, SQL Server, Oracle, MongoDB, Redis, Elasticsearch, CockroachDB, ClickHouse, Couchbase, Neo4j, Snowflake, Trino, and more — 20+ total.

Enterprise Governance#

• Security: IAM authentication integration, structured queries, restricted access to prevent unauthorized operations.
• Observability: Built-in OpenTelemetry metrics and tracing, exportable via --telemetry-otlp to Google Cloud Monitoring and other OTLP-compatible backends.
• Performance: Built-in connection pooling.

Developer Experience#

• Low-Code Integration: Under 10 lines of code to integrate with ADK, LangChain, LlamaIndex and other frameworks.
• Dynamic Reload: Hot-reload of configuration files enabled by default (disable via --disable-reload).
• Interactive Testing UI: Launch Web UI via --ui flag for tool debugging.
• Agent Skills Packaging: skills-generate subcommand converts toolsets into portable Agent Skill packages installable in Gemini CLI.

Architecture#

Middleware control plane sitting between upper-layer orchestration frameworks and lower-layer physical databases, based on standard MCP protocol over HTTP (default port 5000), core endpoint /mcp with toolset-name routing (/mcp/{toolset_name}). Supports custom Source and Tool type extensions (see MCP-TOOLBOX-EXTENSION.md in repository).

Installation#

• NPM: npx @toolbox-sdk/server --config tools.yaml or npx @toolbox-sdk/server --prebuilt=postgres
• Binary: Download platform-specific binaries from Google Cloud Storage
• Docker: Container images available
• Homebrew: brew install supported
• Source: Requires Go toolchain

Multi-Language SDKs#

• Python: toolbox-core, toolbox-langchain (adapts LangChain/LlamaIndex)
• JavaScript/TypeScript: @toolbox-sdk/core (adapts LangChain/Genkit/LlamaIndex/ADK)
• Go: github.com/googleapis/mcp-toolbox-sdk-go (adapts LangChain Go/Genkit Go/OpenAI Go/ADK Go)

Unconfirmed Details#

• Semantic search underlying vector database dependency (e.g., pgvector requirement) not explicitly documented
• NL2SQL implementation mechanism details require further investigation
• Google Antigravity MCP Store public access scope unconfirmed
• Cloudflare integration and Looker tool positioning unconfirmed