This is a really interesting MCP use case.
How are you constraining the AI’s ability to take actions like killing slow queries or creating watches?
Is there a policy layer between the MCP tool invocation and execution, or is it trusting the assistant’s reasoning once the SQL is validated?
Interesting direction. I’m curious how you’re handling schema drift and join discovery when Claude infers relationships.
Are you materializing AI-generated wide tables or regenerating them per query?
I’ve seen LLM-driven schema reasoning struggle once datasets exceed a few hundred columns would love to hear how you’re constraining token usage and maintaining deterministic outputs.