An MCP server that autonomously scans any SQLite database, generates typed CRUD tools for every table it discovers, builds a schema-aware prompt explaining joins/relationships, and enforces Zero-Knowledge security — the LLM never writes raw SQL, only calls pre-validated parameterized templates.
| Feature | Detail |
|---|---|
| Auto-Discovery | Scans sqlite_master at startup; no manual table registration |
| CRUD Generation | Creates get_, create_, update_, delete_ tools per table in memory |
| Schema Prompt | Injects table columns + foreign-key relationships as MCP instructions |
| Zero-Knowledge SQL | Instead of allowing the LLM to write arbitrary SQL (which risks SQL injection and destructive operations), the server generates isolated CRUD tools (Create, Read, Update, Delete) in memory for each table. The LLM interacts with these safe, pre-validated python templates |
| Configurable DB | Point at any SQLite file via DB_PATH env var |
AutoMCP-SQL/
├── server.py # MCP server — scans DB, generates tools, starts server
├── setupdb.py # One-time script to create and seed legacy.db for testing
├── legacy.db # SQLite database (auto-created by setupdb.py)
├── pyproject.toml # Dependencies managed by uv
├── uv.lock # Lockfile — guarantees reproducible installs
└── README.md
| Tool | Why |
|---|---|
| FastMCP | Minimal decorator-based MCP server; zero boilerplate for tool registration |
| sqlite3 (stdlib) | No extra dependency; sufficient for parameterized CRUD on any SQLite file |
| uv | 10-100× faster than pip; lockfile guarantees reproducible installs across machines |
| MCP (Model Context Protocol) | Standard protocol for giving LLMs structured, auditable tool access |
- Python 3.11+
- uv — fast Python package/project manager
- Claude Desktop — to connect the MCP server (optional)
Install uv if you don't have it:
# macOS / Linux
curl -LsSf https://astral.sh/uv/install.sh | sh
# Windows (PowerShell)
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"git clone https://github.com/Mav977/AutoMCP-SQL.git
cd AutoMCP-SQLuv syncThis reads uv.lock and recreates the exact virtual environment.
uv run setupdb.pyThis creates legacy.db in the project root with sample tables (users, orders, products, etc.) so you can test immediately.
Using your own database? Skip this step and point
DB_PATHat your existing.dbfile (see Configuration below).
You can access the configuration file directly from within the app:
- Open Claude Desktop.
- Navigate to Settings.
- Select the Developer tab.
- Click Edit Config to open the
claude_desktop_config.jsonfile in your default text editor.
Replace the path below with the absolute path to your cloned server.py:
Windows:
{
"mcpServers": {
"AutoMCP-SQL": {
"command": "uv",
"args": [
"run",
"--with",
"mcp[cli]",
"mcp",
"run",
"C:\\Users\\YourName\\Desktop\\Projects\\AutoMCP-SQL\\server.py"
],
"env": {
"DB_PATH": "C:\\Users\\YourName\\Desktop\\Projects\\AutoMCP-SQL\\legacy.db"
}
}
}
}macOS / Linux:
{
"mcpServers": {
"AutoMCP-SQL": {
"command": "uv",
"args": [
"run",
"--with",
"mcp[cli]",
"mcp",
"run",
"/home/yourname/projects/AutoMCP-SQL/server.py"
],
"env": {
"DB_PATH": "/home/yourname/projects/AutoMCP-SQL/legacy.db"
}
}
}
}Important: Use double backslashes (
\\) in Windows paths inside JSON.
Fully quit and reopen Claude Desktop. The MCP server starts automatically when Claude launches.
Open a new conversation in Claude Desktop. Click the plus icon (+) in the chat input area, select Connectors, and then click Manage connectors. Here, you will see AutoMCP-SQL listed along with all the tools the MCP has access to (like get_users, create_orders, etc.).
You can also just ask Claude:
"What tables do you have access to?"
Claude will use the injected schema prompt to describe the database structure without querying it.
Set DB_PATH in the config to point at any existing SQLite file:
"env": {
"DB_PATH": "C:\\path\\to\\your\\database.db"
}The server will scan it on startup and auto-generate tools for every table it finds — no code changes needed.
| Environment Variable | Default | Description |
|---|---|---|
DB_PATH |
./legacy.db (next to server.py) |
Path to the SQLite database file |
The server never exposes raw SQL execution to the LLM. Instead:
- On startup,
scan_db()reads the schema and builds a natural-language prompt describing tables and relationships. - For each table, four tool functions are registered with hard-coded SQL templates:
get_<table>→SELECT * FROM <table> WHERE col = ?create_<table>→INSERT INTO <table> (cols) VALUES (?)update_<table>→UPDATE <table> SET col = ? WHERE id = ?delete_<table>→DELETE FROM <table> WHERE id = ?
- All values are passed as parameterized arguments — never string-interpolated into queries.
- Claude can only call these named tools; it has no mechanism to execute arbitrary SQL.
This means even a prompt-injected or jailbroken model cannot run DROP TABLE or exfiltrate data via UNION — the attack surface is limited to the four CRUD templates.
Server doesn't appear in Claude Desktop
- Wait a few seconds; it may take some time to load.
- Double-check the path in
claude_desktop_config.json— it must be the absolute path toserver.py, not a folder. - Make sure
uvis on your system PATH (open a new terminal and runuv --versionto verify). - Fully quit Claude Desktop (system tray on Windows, Cmd+Q on Mac) and reopen it.
DB_PATH does not exist error
Run uv run setupdb.py first to create legacy.db, or update DB_PATH in the config to point at an existing database.
Tools not showing up after config change Claude Desktop only reads the config on launch. Fully restart it after any config edits.