mcp-sql-only 0.1.0

Standalone MCP server exposing a single sql_generator tool for MySQL-backed natural-language-to-SQL

mcp-sql-only

A standalone MCP server that exposes a single sql_generator tool.
Give it a natural-language question and your MySQL connection details — it introspects the schema at runtime, generates a validated SELECT query via Azure OpenAI, and optionally executes it.

How it works

question + MySQL creds
        │
        ▼
  schema introspection  ──►  Azure OpenAI  ──►  SQL validator  ──►  (execute)

1. Connects to your MySQL database and reads the live schema.
2. Sends the schema + question to Azure OpenAI to generate SQL.
3. Validates the SQL (SELECT-only policy, table/column whitelist, no dangerous keywords).
4. Returns the query — or executes it and returns rows.

Required environment variables

Set these before starting the server (in your shell or a .env file in the working directory):

AZURE_OPENAI_API_KEY=your-key
AZURE_OPENAI_ENDPOINT=https://<resource>.openai.azure.com/
AZURE_DEPLOYMENT=your-deployment-name
AZURE_OPENAI_API_VERSION=2024-02-15-preview

Quickstart — run without installing (recommended)

uvx mcp-sql-only

uvx downloads the package from PyPI into a temporary isolated environment and runs it immediately. No global install, no dependency conflicts.

Plug into any MCP-compatible agent

from langchain_mcp_adapters.client import MultiServerMCPClient
from langgraph.prebuilt import create_react_agent

client = MultiServerMCPClient(
    {
        "sql": {
            "transport": "stdio",
            "command": "uvx",
            "args": ["mcp-sql-only"],
        }
    }
)

tools = await client.get_tools()

agent = create_agent(
    model=model,
    tools=tools,
)

Tool — sql_generator

Parameter	Required	Default	Description
question	✅	—	Natural-language question
host	✅	—	MySQL host
username	✅	—	MySQL user
password	✅	—	MySQL password
database_name	✅	—	Target database
port	❌	3306	MySQL port
table	❌	all	Restrict query to a single table
max_retries	❌	0	Retry attempts if SQL validation fails
debug	❌	false	Include internal workflow logs in response

Security

• SELECT-only — DROP, DELETE, UPDATE, INSERT, TRUNCATE and other destructive statements are blocked before execution.
• Schema whitelist — generated SQL is validated against the live schema; unknown tables and columns are rejected.
• No credential logging — passwords are masked as *** in all server logs.
• Credentials are never stored — connection details are passed per-call and held only for the duration of that request.

Project structure

server.py           ← MCP server entry point
sql_engine/
  generator.py      ← LLM prompt + SQL generation
  validator.py      ← SELECT-only + whitelist validation
  llm_providers.py  ← Azure OpenAI provider
  tool.py           ← orchestrates generate → validate → execute
  models.py         ← Pydantic input/output contracts
pyproject.toml      ← package metadata + entry point