I hit the same roadblock with MCP. If you work with data, LLM becomes a very expensive pipe with an added risk of hallucinations. It’s better to simply connect it to a Python environment enriched with integrations you need.