Google MCP Toolbox MCP Server

MCP Toolbox for Databases (formerly genai-toolbox) is an open source Model Context Protocol server maintained by Google under the googleapis GitHub org. It bridges AI agents, IDEs, and applications to enterprise databases, exposing both ready-to-use prebuilt tools for schema exploration and SQL execution, and a framework for defining custom, structured tools via a tools.yaml config file.

The toolbox ships with prebuilt configurations for a wide range of databases. Google Cloud sources include AlloyDB, BigQuery, Cloud SQL (PostgreSQL, MySQL, SQL Server), Spanner (GoogleSQL and PostgreSQL dialects), Firestore, Bigtable, and Knowledge Catalog. It also supports standalone PostgreSQL, MySQL, SQL Server, SQLite, Oracle, MongoDB, Redis, Elasticsearch, Neo4j, Snowflake, ClickHouse, Couchbase, CockroachDB, TiDB, OceanBase, and Trino. Beyond raw database access, prebuilt configs are available for Looker, Cloud Storage, Cloud Monitoring, Cloud Logging, Dataproc, Serverless Spark, and Gemini Data Analytics.

Notable platform features include connection pooling, integrated IAM authentication, OpenTelemetry-based observability, dynamic tool reloading during development, and an interactive --ui mode for testing tools. SDKs are published for Python, JavaScript/TypeScript, Go, and Java, with native integrations for LangChain, LlamaIndex, Genkit, and Google's Agent Development Kit (ADK).

Tools

Prerequisites

• Credentials for the database you want to expose (connection string, service account, or IAM principal).
• One of: npx, Docker, Homebrew, a downloaded binary, or Go installed locally.

Option A: Run with npx (quickest)

Add the server to your MCP client config (Claude Desktop, Cursor, Gemini CLI, etc.):

{
  "mcpServers": {
    "toolbox-postgres": {
      "command": "npx",
      "args": [
        "-y",
        "@toolbox-sdk/server",
        "--prebuilt=postgres",
        "--stdio"
      ],
      "env": {
        "POSTGRES_HOST": "127.0.0.1",
        "POSTGRES_PORT": "5432",
        "POSTGRES_DATABASE": "my_db",
        "POSTGRES_USER": "my_user",
        "POSTGRES_PASSWORD": "my_password"
      }
    }
  }
}

Swap --prebuilt=postgres for bigquery, alloydb, cloud-sql-pg, cloud-sql-mysql, spanner, firestore, bigtable, mysql, etc.

Option B: Install the binary

# Homebrew
brew install mcp-toolbox

# Or via Go
go install github.com/googleapis/mcp-toolbox@latest

# Or via Docker
docker pull us-central1-docker.pkg.dev/database-toolbox/toolbox/toolbox:latest

Run with a custom tools.yaml:

mcp-toolbox --tools-file tools.yaml --port 5000

Then point your MCP client at the HTTP endpoint:

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

Example tools.yaml

sources:
  my-pg-source:
    kind: postgres
    host: 127.0.0.1
    port: 5432
    database: toolbox_db
    user: toolbox_user
    password: ${POSTGRES_PASSWORD}

tools:
  search-hotels-by-name:
    kind: postgres-sql
    source: my-pg-source
    description: Search hotels by name.
    parameters:
      - name: name
        type: string
        description: Hotel name to search for
    statement: SELECT * FROM hotels WHERE name ILIKE '%' || $1 || '%';

• Expose a read-only BigQuery or Cloud SQL instance to Claude Code or Gemini CLI so an engineer can ask natural language questions about production data without writing SQL by hand.
• Give an agent secure, parameterized access to specific queries (e.g. search-orders-by-customer) using tools.yaml instead of granting raw SQL execution.
• Let a Firestore-backed product assistant fetch documents and run schema introspection during conversations.
• Run a centralized Toolbox HTTP server inside a VPC so multiple ADK or LangChain agents share connection pooling and IAM auth against AlloyDB or Spanner.
• Connect an LLM-based BI assistant to Looker or Gemini Data Analytics for conversational reporting via the prebuilt config.

• "List all tables in my BigQuery analytics dataset and show me the row count for each."
• "Using the Cloud SQL Postgres source, find the top 10 customers by revenue last quarter."
• "Run list_tables then describe the schema of the orders table in Spanner."
• "Call the search-hotels-by-name tool with name='Marriott' and summarize the results."
• "Query Firestore for all documents in the users collection where signup_source = 'paid'."

• Officially maintained by Google under the googleapis org, with active releases and first-class support for Google Cloud databases.
• Broad database coverage: AlloyDB, BigQuery, Cloud SQL, Spanner, Firestore, Bigtable, plus Postgres, MySQL, MongoDB, Neo4j, Snowflake, and others.
• Production-oriented: connection pooling, IAM/auth integration, OpenTelemetry observability, and an HTTP transport for shared multi-agent deployments.
• Custom tools.yaml lets teams expose narrowly scoped, parameterized queries instead of raw SQL, which is safer than generic execute_sql for production agents.

• The tools.yaml configuration model has a learning curve compared to drop-in MCP servers that need no config.
• Prebuilt generic tools like execute_sql give agents broad access; locking things down requires writing custom tools or running read-only credentials.
• Some advanced features (semantic search, ADK integrations) require additional SDKs or Google Cloud services to be useful.

• Supabase MCP Server for Postgres-backed Supabase projects.
• PostgreSQL MCP server from the official modelcontextprotocol/servers repo for simple read-only Postgres access.
• MongoDB MCP Server for MongoDB-specific workflows.