db-connect-mcp - Multi-Database MCP Server

<!-- mcp-name: io.github.yugui923/db-connect-mcp -->

A read-only MCP (Model Context Protocol) server for exploratory data analysis across multiple database systems. This server provides safe, read-only access to PostgreSQL, MySQL, and ClickHouse databases with comprehensive analysis capabilities.

Demo

db-connect-mcp demo

Quick Start

1. Install:

   pip install db-connect-mcp

1. Add to Claude Desktop claude_desktop_config.json:

   {
     "mcpServers": {
       "db-connect": {
         "command": "python",
         "args": ["-m", "db_connect_mcp"],
         "env": {
           "DATABASE_URL": "postgresql://user:pass@localhost:5432/mydb"
         }
       }
     }
   }

1. Restart Claude Desktop and start querying your database!

Note: Using python -m db_connect_mcp ensures the command works even if Python's Scripts directory isn't in your PATH. Existing mcpServers configurations remain valid: stdio is still the default. The optional remote Streamable HTTP mode serves both older initialize-based clients and 2026-07-28 session-free MCP clients at /mcp without changing the database or SSH configuration. Existing HTTP deployments must review forwarded Host/Origin values: the upgrade enables transport validation and may require MCP_ALLOWED_HOSTS / MCP_ALLOWED_ORIGINS.

Features

🗄️ Multi-Database Support

• PostgreSQL - Full support with advanced metadata and statistics
• MySQL - Complete support for MySQL and MariaDB databases
• ClickHouse - Support for analytical workloads and columnar storage

🔍 Database Exploration

• List schemas - View all schemas in the database
• List tables - See all tables with metadata (size, row counts, comments)
• Describe tables - Get detailed column information, indexes, and constraints
• View relationships - Understand foreign key relationships between tables

📊 Data Analysis

• Column profiling - Statistical analysis of column data
• Basic statistics (count, unique values, nulls)
• Numeric statistics (mean, median, std dev, quartiles)
• Value frequency distribution
• Cardinality analysis
• Data sampling - Preview table data with configurable limits
• Custom queries - Execute read-only SQL queries safely
• Object search - Find schemas, tables, views, columns, and indexes without loading the full catalog
• Query plans - Inspect estimated plans or opt into EXPLAIN ANALYZE where supported
• Hypothetical indexes (PostgreSQL, opt-in) - Compare HypoPG planner estimates without creating a real index. Requires an administrator-installed HypoPG extension, DB_HYPOPG_ENABLED=true on the server, and opt_in=true on each compare_hypothetical_index call. Provide a single simple SELECT, schema, table, and 1–4 column names; comments, CTEs, functions, literals, parameters, and multiple statements are rejected. The tool uses one read-only connection and EXPLAIN without ANALYZE, caps planning/lock waits at 5s/1s, and rolls back session-local changes. The response reports hypothetical_index_used by matching the HypoPG index name in the resulting plan; an unused candidate may be irrelevant or simply not chosen by the planner. The candidate requires schema USAGE and table SELECT privileges. These estimates are not measured speedups.

🔒 Safety Features

• Read-only enforcement - All connections are read-only at multiple levels
• Query validation - Only SELECT and WITH queries are allowed
• Automatic limits - Queries are automatically limited to prevent large result sets
• Connection string safety - Automatically adds read-only parameters
• Database-specific safety - Each adapter implements appropriate safety measures

🔭 Observability

db-connect-mcp inherits the MCP SDK's built-in OpenTelemetry server instrumentation. The API is a no-op until the launching process configures an SDK and exporter. Review exporter sampling and redaction before production use, because database identifiers and error details may be sensitive.

💡 Best Practices

Tip: db-connect-mcp works best with databases that have proper comments on tables and columns. When your database includes descriptive comments, the MCP server can provide richer context to AI assistants, leading to better understanding of your data model and more accurate query suggestions.

Adding comments in PostgreSQL:

COMMENT ON TABLE users IS 'Registered user accounts with profile information';
COMMENT ON COLUMN users.email IS 'Primary email address, used for authentication';
COMMENT ON COLUMN users.is_verified IS 'Whether email has been verified via confirmation link';

Adding comments in MySQL:

ALTER TABLE users COMMENT = 'Registered user accounts with profile information';
ALTER TABLE users MODIFY COLUMN email VARCHAR(255) COMMENT 'Primary email address, used for authentication';

The server automatically retrieves and displays these comments when describing tables, helping AI assistants understand the purpose and semantics of your data.

🔐 SSH Tunnel Support

• Secure remote access - Connect to databases behind firewalls via SSH tunnels
• Automatic tunnel management - Tunnel lifecycle handled transparently (start, health check, restart, cleanup)
• Reliable native forwarding - Paramiko SSHClient transport with target preflight and stable-port recovery
• Flexible authentication - Password or private key based SSH authentication
• Any database type - Works with PostgreSQL, MySQL, and ClickHouse through the same tunnel

See the SSH Tunnel Guide for configuration details.

Installation

Prerequisites

• Python 3.10 or higher
• A database: PostgreSQL (9.6+), MySQL/MariaDB (5.7+/10.2+), or ClickHouse

Install via pip

pip install db-connect-mcp

That's it! The package is now ready to use.

For developers: See Development Guide for setting up a development environment.

Configuration

Create a .env file with your database connection string:

DATABASE_URL=your_database_connection_string_here

The server automatically detects the database type and adds appropriate read-only parameters.

For PostgreSQL, set ENABLE_PG_STAT_STATEMENTS=true to advertise the optional get_top_queries tool. It reads bounded cumulative statistics for the current user/database if pg_stat_statements was installed and preloaded by the database administrator. It does not install extensions. Query text is omitted unless explicitly requested and can contain sensitive information. See the top-query report contract for bounds and statuses.

Connection String Examples

The server now provides more flexible and secure URL handling:

• Automatic driver detection: Async drivers are automatically added if not specified
• JDBC URL support: JDBC prefixes are automatically handled
• jdbc:postgresql://... → postgresql+asyncpg://...
• jdbc:mysql://... → mysql+aiomysql://...
• Works with all dialect variations (e.g., jdbc:postgres://, jdbc:mariadb://)
• Database dialect variations: Common variations are automatically normalized
• PostgreSQL: postgresql, postgres, pg, psql, pgsql
• MySQL/MariaDB: mysql, mariadb, maria
• ClickHouse: clickhouse, ch, click
• Allowlist-based parameter filtering: Only known-safe parameters are preserved
• Database-specific parameters: Each database type has its own set of supported parameters
• Robust parsing: Handles various URL formats gracefully

PostgreSQL:

# Simple URL (driver automatically added)
DATABASE_URL=postgresql://user:password@localhost:5432/mydb

# Common variations (all normalized to postgresql+asyncpg)
DATABASE_URL=postgres://user:pass@host:5432/db  # Heroku, AWS RDS style
DATABASE_URL=pg://user:pass@host:5432/db         # Short form
DATABASE_URL=psql://user:pass@host:5432/db       # CLI style

# JDBC URLs (automatically converted)
DATABASE_URL=jdbc:postgresql://user:pass@host:5432/db  # From Java apps
DATABASE_URL=jdbc:postgres://user:pass@host:5432/db    # JDBC with variant

# With explicit async driver
DATABASE_URL=postgresql+asyncpg://user:pass@host:5432/db

# With supported parameters (see list below)
DATABASE_URL=postgres://user:pass@host:5432/db?application_name=myapp&connect_timeout=10

Supported PostgreSQL Parameters:

• application_name - Identifies your app in pg_stat_activity (useful for monitoring)
• connect_timeout - Connection timeout in seconds
• command_timeout - Default timeout for operations
• ssl / sslmode - SSL connection requirements (automatically converted for asyncpg compatibility)
• server_settings - Server settings dictionary
• options - Command-line options to send to server
• Performance tuning: prepared_statement_cache_size, max_cached_statement_lifetime, etc.

MySQL/MariaDB:

# Simple URL (driver automatically added)
DATABASE_URL=mysql://root:password@localhost:3306/mydb

# MariaDB URLs (normalized to mysql+aiomysql)
DATABASE_URL=mariadb://user:pass@host:3306/db    # MariaDB style
DATABASE_URL=maria://user:pass@host:3306/db      # Short form

# JDBC URLs (automatically converted)
DATABASE_URL=jdbc:mysql://user:pass@host:3306/db     # From Java apps
DATABASE_URL=jdbc:mariadb://user:pass@host:3306/db   # JDBC MariaDB

# With explicit async driver
DATABASE_URL=mysql+aiomysql://user:pass@host:3306/db

# With charset (critical for proper Unicode support)
DATABASE_URL=mariadb://user:[email protected]:3306/db?charset=utf8mb4

Supported MySQL Parameters:

• charset - Character encoding (e.g., utf8mb4) - critical for data integrity
• use_unicode - Enable Unicode support
• connect_timeout, read_timeout, write_timeout - Various timeouts
• autocommit - Transaction autocommit mode
• init_command - Initial SQL command to run
• sql_mode - SQL mode settings
• time_zone - Time zone setting

ClickHouse:

# Simple URL (driver automatically added)
DATABASE_URL=clickhouse://default:@localhost:9000/default

# Short forms (normalized to clickhouse+asynch)
DATABASE_URL=ch://user:pass@host:9000/db         # Short form
DATABASE_URL=click://user:pass@host:9000/db      # Alternative

# JDBC URLs (automatically converted)
DATABASE_URL=jdbc:clickhouse://user:pass@host:9000/db  # From Java apps
DATABASE_URL=jdbc:ch://user:pass@host:9000/db         # JDBC with short form

# With explicit async driver
DATABASE_URL=clickhouse+asynch://user:pass@host:9000/db

# With performance settings
DATABASE_URL=ch://user:pass@host:9000/db?timeout=60&max_threads=4

Supported ClickHouse Parameters:

• database - Default database selection
• timeout, connect_timeout, send_receive_timeout - Various timeouts
• compress, compression - Enable compression
• max_block_size, max_threads - Performance tuning

Note:

• SSL parameters (ssl, sslmode) are automatically converted to the correct format for asyncpg
• Certificate file parameters (sslcert, sslkey, sslrootcert) are filtered out as they can cause compatibility issues
• Only parameters known to work with async drivers are preserved

Usage

Running the Server

# Run the server (works everywhere, no PATH configuration needed)
python -m db_connect_mcp

# With environment variable
DATABASE_URL="postgresql://user:pass@host:5432/db" python -m db_connect_mcp

Note: Using python -m db_connect_mcp works regardless of whether Python's Scripts directory is in your PATH.

Using with Claude Code

Add the MCP server to your project's .mcp.json:

claude mcp add --transport stdio db-connect --scope project \
  --env DATABASE_URL=postgresql://user:pass@host:5432/db \
  -- python -m db_connect_mcp

Or manually create .mcp.json in your project root. Below are examples for each supported database:

PostgreSQL:

{
  "mcpServers": {
    "db-connect-mcp": {
      "command": "python",
      "args": ["-m", "db_connect_mcp"],
      "env": {
        "DATABASE_URL": "postgresql+asyncpg://user:pass@host:5432/mydb"
      }
    }
  }
}

MySQL:

{
  "mcpServers": {
    "db-connect-mcp": {
      "command": "python",
      "args": ["-m", "db_connect_mcp"],
      "env": {
        "DATABASE_URL": "mysql+aiomysql://user:pass@host:3306/mydb"
      }
    }
  }
}

ClickHouse:

{
  "mcpServers": {
    "db-connect-mcp": {
      "command": "python",
      "args": ["-m", "db_connect_mcp"],
      "env": {
        "DATABASE_URL": "clickhouse+asynch://default:@host:9000/default"
      }
    }
  }
}

PostgreSQL via SSH tunnel (database behind a firewall, reachable only through a bastion host):

{
  "mcpServers": {
    "db-connect-mcp": {
      "command": "python",
      "args": ["-m", "db_connect_mcp"],
      "env": {
        "DATABASE_URL": "postgresql+asyncpg://user:pass@db-internal:5432/mydb",
        "SSH_HOST": "bastion.example.com",
        "SSH_PORT": "22",
        "SSH_USERNAME": "deployer",
        "SSH_PRIVATE_KEY_PATH": "/home/user/.ssh/id_rsa"
      }
    }
  }
}

MySQL via SSH tunnel:

{
  "mcpServers": {
    "db-connect-mcp": {
      "command": "python",
      "args": ["-m", "db_connect_mcp"],
      "env": {
        "DATABASE_URL": "mysql+aiomysql://user:pass@db-internal:3306/mydb",
        "SSH_HOST": "bastion.example.com",
        "SSH_PORT": "22",
        "SSH_USERNAME": "deployer",
        "SSH_PASSWORD": "secret"
      }
    }
  }
}

Multiple databases (each MCP server instance connects to one database):

{
  "mcpServers": {
    "postgres-prod": {
      "command": "python",
      "args": ["-m", "db_connect_mcp"],
      "env": {
        "DATABASE_URL": "postgresql+asyncpg://user:pass@pg-host:5432/prod"
      }
    },
    "mysql-analytics": {
      "command": "python",
      "args": ["-m", "db_connect_mcp"],
      "env": {
        "DATABASE_URL": "mysql+aiomysql://user:pass@mysql-host:3306/analytics"
      }
    }
  }
}

After creating .mcp.json, restart Claude Code and verify with /mcp. You should see db-connect-mcp listed with all available tools.

Tip: Instead of SSH_PRIVATE_KEY_PATH, you can use SSH_PRIVATE_KEY to pass the private key content directly as a string (raw PEM or base64-encoded PEM). This is useful in CI/CD or cloud environments where mounting key files is impractical.

See the SSH Tunnel Guide for full tunnel configuration reference.

Using with Claude Desktop

Add the server to your Claude Desktop configuration (claude_desktop_config.json):

{
  "mcpServers": {
    "db-connect": {
      "command": "python",
      "args": ["-m", "db_connect_mcp"],
      "env": {
        "DATABASE_URL": "postgresql+asyncpg://user:pass@host:5432/db"
      }
    }
  }
}

The same database URL formats and SSH tunnel environment variables shown in the Claude Code examples above work identically with Claude Desktop.

For development: See Development Guide for running from source with uv.

Database Feature Support

MCP Resources

Modern MCP clients can discover database context as private, cache-aware JSON resources in addition to calling tools:

• db-connect://database — database identity, dialect, and capabilities
• db-connect://schema/{schema} — schema counts and metadata
• db-connect://table/{schema}/{table} — columns, indexes, constraints, and comments

The schema and table forms are also advertised as resource templates for direct access when the identifier is already known.

Resource catalogs are URI-sorted and cursor-paginated in pages of 100. Cursors are tied to a catalog snapshot; if schemas or tables change between pages, the server asks the client to restart pagination instead of returning an inconsistent traversal.

Available Tools

All tools publish JSON Schema input and output contracts, read-only behavior annotations, and machine-readable structured results. The same result remains available as JSON text for clients that do not yet consume MCP structured content. Structured list results use an items envelope while their legacy text form remains a JSON array.

get_database_info

Get database metadata, including the dialect, version, connection details, read-only status, and capabilities.

list_schemas

List all schemas in the database.

list_tables

List all tables in a schema with metadata.

• Parameters:
• schema (optional): Schema name (default: "public")

describe_table

Get detailed information about a table.

• Parameters:
• table: Name of the table
• schema (optional): Schema name (default: "public")

analyze_column

Analyze a column with statistics and distribution.

• Parameters:
• table: Name of the table
• column: Name of the column
• schema (optional): Schema name (default: "public")

sample_data

Get a sample of data from a table.

• Parameters:
• table: Name of the table
• schema (optional): Schema name (default: "public")
• limit (optional): Number of rows (default: 100, max: 1000)

execute_query

Execute a read-only SQL query.

• Parameters:
• query: SQL query (must be SELECT or WITH)
• limit (optional): Maximum rows (default: 1000, max: 10000)

get_table_relationships

Get foreign key relationships for a table.

• Parameters:
• table: Name of the table
• schema (optional): Schema name (default: "public")

explain_query

Get a database-specific query execution plan.

• Parameters:
• query: SQL query to explain
• analyze (optional): Execute the query and include actual runtime statistics (default: false)

search_objects

Search schemas, tables, views, columns, and indexes with progressive detail.

• Parameters:
• pattern: SQL LIKE pattern, such as %user%
• object_types (optional): Object types to include
• detail_level (optional): names, summary, or full (default: summary)
• schema (optional): Restrict the search to a schema
• table (optional): Restrict column and index searches to a table
• limit (optional): Maximum matches (default: 100, max: 1000)

Example Usage

Once configured, you can use the server with any compatible MCP client:

"Can you analyze my database and tell me about the table structure?"

"Show me the relationships between tables in the public schema"

"What's the distribution of values in the users.created_at column?"

"Give me a sample of data from the orders table"

"Run this query: SELECT COUNT(*) FROM users WHERE created_at > '2024-01-01'"

Database-Specific Examples

Working with PostgreSQL:

"List all schemas except system ones"
"Show me the foreign key relationships in the sales schema"
"Analyze the performance of indexes on the products table"

Working with MySQL:

"What storage engines are being used in my database?"
"Show me all tables in the information_schema"
"Analyze the customer_orders table structure"

Working with ClickHouse:

"Show me the partitions for the events table"
"What's the compression ratio for the analytics.clicks table?"
"Sample 1000 rows from the metrics table"

Safety and Security

• Read-only intent: Write statements are rejected, and PostgreSQL and MySQL connections use read-only transactions. This is not a SQL sandbox: SELECT may call functions with external or session-level effects. MySQL read-only sessions are not pooled across requests; direct named-lock calls and variable assignments are rejected in execute_query. Use a least-privilege database role on every engine, especially ClickHouse, where SELECT may invoke external table functions.
• Query limits: execute_query bounds returned rows and serialized response size, not total database work or peak process memory. ClickHouse's synchronous driver may buffer a full result before the application checks it. Use database-side execution, memory, and access limits. See health and query safety notes.

Development

For detailed development setup, testing, and contribution guidelines, see the Development Guide.

Project Structure

db-connect-mcp/
├── src/
│   └── db_connect_mcp/
│       ├── adapters/         # Database-specific adapters
│       │   ├── __init__.py
│       │   ├── base.py      # Base adapter interface
│       │   ├── postgresql.py # PostgreSQL adapter
│       │   ├── mysql.py     # MySQL adapter
│       │   └── clickhouse.py # ClickHouse adapter
│       ├── core/            # Core functionality
│       │   ├── __init__.py
│       │   ├── connection.py # Database connection management
│       │   ├── executor.py  # Query execution
│       │   ├── inspector.py # Metadata inspection
│       │   ├── analyzer.py  # Statistical analysis
│       │   └── tunnel.py   # SSH tunnel management
│       ├── models/          # Data models
│       │   ├── __init__.py
│       │   ├── capabilities.py # Database capabilities
│       │   ├── config.py    # Configuration models
│       │   ├── database.py  # Database models
│       │   ├── query.py     # Query models
│       │   ├── statistics.py # Statistics models
│       │   └── table.py     # Table metadata models
│       ├── __init__.py
│       ├── __main__.py      # Module entry point
│       └── server.py        # Main MCP server implementation
├── tests/
│   ├── unit/            # Unit tests (mocked)
│   ├── module/          # Module tests (single component + DB)
│   ├── integration/     # Integration tests (full stack)
│   └── conftest.py      # Shared fixtures
├── .env.example         # Example environment configuration
├── pyproject.toml      # Project dependencies and console scripts
└── README.md          # This file

Architecture

The server uses an adapter pattern to support multiple database systems:

• Adapters: Each database type has its own adapter that implements database-specific functionality
• Core: Shared functionality for connection management, query execution, and metadata inspection
• Models: Pydantic models for type safety and validation
• Server: MCP server implementation that routes requests to appropriate components

Running Tests

# Start local test database (PostgreSQL 17 with sample data)
cd tests/docker && docker-compose up -d && cd ../..

# Run all tests in parallel (preferred - 6 workers)
uv run pytest -n 6

# Run specific test modules
uv run pytest tests/module/test_inspector.py -v -n 6
uv run pytest tests/integration/ -v -n 6

# Stop test database
cd tests/docker && docker-compose down && cd ../..

# Reset database (clean slate with fresh data)
cd tests/docker && docker-compose down -v && docker-compose up -d && cd ../..

Local Test Database:

• PostgreSQL 17 with 50K+ rows of sample data across 7 tables
• Automatically initialized via Docker Compose
• No cloud database or .env configuration required
• See Docker Setup for details

See the Development Guide and Testing Guide for detailed testing instructions.

Troubleshooting

Connection Issues

• Verify your DATABASE_URL is correct and includes the appropriate driver
• Check network connectivity to the database
• Ensure the database user has appropriate read permissions
• For PostgreSQL: Check if SSL is required (?ssl=require)
• For MySQL: Verify charset settings (?charset=utf8mb4)
• For ClickHouse: Check port (default is 9000 for native, 8123 for HTTP)

Database-Specific Issues

PostgreSQL:

• Ensure asyncpg driver is specified for async operations
• SSL certificates may be required for cloud databases

MySQL/MariaDB:

• Use aiomysql driver for async support
• Check MySQL version compatibility (5.7+ or MariaDB 10.2+)
• Verify charset and collation settings

ClickHouse:

• Use asynch driver for async operations
• Note that ClickHouse has limited support for foreign keys and constraints
• Some statistical functions may not be available

Permission Errors

• The database user needs at least SELECT permissions on the schemas/tables you want to analyze
• Some statistical functions may require additional permissions
• ClickHouse may require specific permissions for system tables

Large Result Sets

• Use the limit parameter to control result size
• The server automatically limits results to prevent memory issues
• For large analyses, consider using more specific queries

Author

Created by Yuri Gui.

Contributing

Contributions are welcome! The server is designed to be read-only and safe by default. Any new features should maintain these safety guarantees.

License

MIT License - See LICENSE file for details