I created an MCP server aimed at giving LLM agents safe access to a database. It's part of the broader Polymr platform (https://polymr-platform.github.io/).

I generally allow the LLM to perform selects as it wishes to establish context, but once it comes down to modifying data, especially in a sensitive environment, I want to retain absolute control and understanding of what it's trying to do.

That's why I built mcp-sql.

Some of the features:

- DML Previews: Queries can be run in preview mode where the server executes the transaction and immediately rolls it back. This lets you see exactly what will happen if you allow the change.

- Dynamic Permission Prompts: The server dynamically introspects SQL to determine required permissions. If an agent tries to execute a query it doesn't have auto-approval for, it prompts the user. You retain full control over the agent's autonomy level.

- Environment Routing: You can use policies to dynamically switch the connection a query is run on, allowing you to easily switch between environments.

I'd love feedback, especially from people using MCP servers with production databases.