mcp-sql gives AI agents safe, human-controlled database access

`mcp-sql` is an open-source MCP server posted by its-a-new-world on Hacker News, designed to give LLM agents safe access to databases as part of the Polymr platform. The core design philosophy separates read access from write access: agents are permitted to run SELECT queries freely to establish context, while any data-modifying operations require explicit human oversight.

DML Previews execute a transaction and immediately roll it back, letting the user see exactly what a proposed change would do before approving it.

The server ships with three headline features. DML Previews execute a transaction and immediately roll it back, letting the user see exactly what a proposed change would do before approving it. Dynamic Permission Prompts introspect the SQL being executed to determine what permissions are required, and if the agent attempts a query it doesn't have auto-approval for, the server prompts the user — keeping the human in control of the agent's autonomy level. Environment Routing allows policies to dynamically switch the database connection a query runs against, making it straightforward to route between environments such as staging and production.

Key facts

1. 01mcp-sql is an MCP server for giving LLM agents safe, controlled access to databases
2. 02Part of the broader Polymr platform (polymr-platform.github.io)
3. 03Agents can run SELECT queries freely; data-modifying (DML) operations require human approval
4. 04DML Previews execute a transaction and immediately roll it back to show the effect without committing
5. 05Dynamic Permission Prompts introspect SQL to determine required permissions and prompt the user if auto-approval is absent
6. 06Environment Routing uses policies to dynamically switch the database connection a query runs on

Topics