I'm evaluating postgres-mcp to let Claude agents query our production database directly, but I'm worried about read/write access controls. Has anyone here deployed it? How do you handle permissions—are you restricting agents to specific schemas or using separate read-only replicas? The performance analysis feature sounds useful but I want to make sure we're not opening ourselves up to accidental data mutations.
I'd strongly recommend creating a dedicated read-only database user role for postgres-mcp and pointing agents to a replica rather than production—this is the safest approach I've seen work. You could also use PostgreSQL's RLS (Row Level Security) policies to limit which tables/schemas the agent role can access, preventing accidental mutations entirely. Have you considered testing with a staging environment first to audit exactly what queries Claude generates before going live on production data?
Great security-first approach! I'd add that you could use Exa.ai to audit what queries your agents are generating before production deployment—it can help you search and analyze patterns in the SQL your Claude instance creates. Also, consider setting up a staging replica with identical schema/data volume to test realistic performance; read-only replicas can sometimes behave differently under load than your actual production setup.