Database add-on

The database MCP server gives Claude three tools to work with your own SQL database: list the tables, describe a table's columns, and run a query. It supports MySQL and MariaDB-compatible servers as well as PostgreSQL. You ask "how many accounts signed up per month this year?", Claude looks at the schema, writes the SQL, runs it and explains the result.

This add-on needs mcp-core, the self-hosted MCP server. mcp-core runs in Docker on your infrastructure and opens the database connection from there. Your database credentials stay in your own .env and never leave your server.

How does Claude use the database MCP server?

Claude works in three steps, the same way a developer new to the codebase would:

1. database_list_tables to see what exists.
2. database_describe_table to read column names, data types, nullability and defaults.
3. database_query to run one SQL statement and get the rows back.

Because Claude reads the schema first, you can ask questions in plain language and let it work out the joins. The SQL is visible in each tool call, so you can check it or reuse it elsewhere.

Is it safe to point Claude at a production database?

It is safe when you combine the built-in guardrails with a restricted database user. The tool starts in read-only mode: only SELECT reaches the database, and INSERT, UPDATE and DELETE are rejected with a clear error. CREATE, DROP, ALTER, TRUNCATE, GRANT and REVOKE are blocked in every mode. Only one statement per call is accepted.

Be clear about what that check is. It inspects the SQL text, so it is a safety net and not a permission system. The real enforcement belongs in the database itself: connect the database MCP server with a user that has SELECT only, and only on the tables or views Claude should see. The database setup guide lists every variable and what it does.

Which limits can you set on queries?

Two variables bound every query that the database MCP server runs. DB_MAX_ROWS caps the rows returned to Claude (default 200), and Claude gets a note when a result was cut off so it can narrow the query. DB_QUERY_TIMEOUT_MS stops slow queries (default 10,000 ms). A forgotten WHERE clause on a large table returns a capped answer instead of tying up your server.

Which settings does the database MCP server need?

You configure it with standard connection details in the .env next to your docker-compose.yml:

DB_TYPE=postgres
DB_HOST=db.internal.example.com
DB_PORT=5432
DB_NAME=app_production
DB_USER=claude_readonly
DB_PASSWORD=change-me
DB_SSL=true
DB_QUERY_TIMEOUT_MS=10000
DB_MAX_ROWS=200

Use DB_TYPE=mysql for MySQL or MariaDB. DB_PORT, DB_SSL and the two limits are optional. Run ./mcp-update.sh after editing the file. If you buy the add-on on an existing install, docker exec mcp-core npm run refresh-license loads it without waiting for the next license check.

When would you allow write queries?

Give a seat read-write mode when Claude should also run INSERT, UPDATE or DELETE, for example to fix a batch of bad records you reviewed together. You set the mode per seat when you add it in My Account → MCP Self-hosted: a developer gets read-write, everyone else on your team keeps read-only access to the same database MCP server. DDL stays blocked either way.

Who is it for?

The database MCP server suits developers who want answers from their application data without writing every query by hand. Agencies that support client applications and need quick checks during an incident. Product and finance people who need numbers, with a developer setting up the read-only user once.

If the database behind your shop is WooCommerce, also look at the WooCommerce MCP server: it uses the REST API and returns orders and products in a ready-made format, while SQL is better for custom reports. The guide on connecting Claude to a database explains the approach for a less technical audience.