MySQL Ops MCP

English | 简体中文

A MySQL database MCP Server. Let AI tools like Claude Code / Cursor / Codex talk to your MySQL databases in natural language — browse databases and tables, inspect schemas, run queries, and perform CRUD operations, with destructive operations (DELETE / DROP) gated behind explicit user approval.

Features

🗄️ Schema browsing: list databases, list tables, describe table structure, full table info (indexes, DDL, stats)
🔍 Queries: run SELECT queries against any database
✏️ CRUD: create tables, insert, update — with SQL-statement validation per tool
🔐 Approval-gated destructive ops: delete_data and drop_table always require explicit user confirmation in the client
🔌 Zero install: runs via npx, configured entirely through environment variables

Installation

Requires Node.js 18+. No clone or build needed — npx fetches and runs the published npm package on first use (see the config below).

To build from source instead (e.g. for development):

bash

git clone https://github.com/GT-dinuo/mysql-ops-mcp.git
cd mysql-ops-mcp
npm install
npm run build

Build output goes to dist/; the entry point is dist/index.js.

Configuration

In each project where you want to use this tool, create (or append to) .mcp.json:

config.json

{
  "mcpServers": {
    "mysql": {
      "command": "npx",
      "args": ["-y", "mysql-ops-mcp"],
      "env": {
        "DB_HOST": "localhost",
        "DB_PORT": "3306",
        "DB_USER": "root",
        "DB_PASSWORD": "your_password",
        "DB_DATABASE": "your_database"
      }
    }
  }
}

If you built from source, point the command at the local build instead: "command": "node", "args": ["/absolute/path/to/mysql-ops-mcp/dist/index.js"].

Environment Variables

Variable	Required	Description
`DB_HOST`	No	MySQL host, default `localhost`
`DB_PORT`	No	MySQL port, default `3306`
`DB_USER`	No	MySQL user, default `root`
`DB_PASSWORD`	No	MySQL password
`DB_DATABASE`	No	Default database; if omitted, pass `database` per tool call

Usage

Once configured, just ask in natural language in your AI tool:

Code

List all databases
Show the structure of the users table
Query the 10 most recent orders
Get full info for the users table — indexes and DDL included
Insert a test record into the users table
Delete the test record you just inserted   ← the client will ask for your approval first

Destructive operations (delete_data, drop_table) are always confirmed by the MCP client before execution — nothing irreversible happens silently.

Tool List

Tool	Description
`list_databases`	List all databases
`list_tables`	List tables in a database
`describe_table`	Show table columns and types
`get_table_info`	Full table details: columns, indexes, DDL, stats
`query`	Run a SELECT query
`create_table`	Create a table (CREATE TABLE only)
`insert`	Insert rows (INSERT only)
`update`	Update rows (UPDATE only)
`delete_data`	Delete rows — requires user approval
`drop_table`	Drop a table — requires user approval
`execute_sql`	Run any custom SQL statement

Security

Destructive operations require approval: delete_data and drop_table are gated by the MCP client's tool-call confirmation — approve or reject each one.
Statement validation: each write tool only accepts its own SQL verb (e.g. query rejects anything but SELECT).
Use a least-privilege account: don't connect as root in production — create a MySQL user scoped to the databases the AI may touch.
Never commit credentials: .env / .mcp.json with real passwords must stay out of version control (.gitignore already covers .env).

Testing the Connection

After configuring .env (copy from env.example), you can verify connectivity before wiring up the client: