Hey, speaking of database tables, how does your agent communicate with them? I managed to get mssql-python mcp working after huge struggles, and it was working just fine. But after 1.5.x I noticed many times it wants to know what data table has, it says the tool did not return any actual results.
And also, I wanted to do some little bit more advanced migration from one structure to another, and while agent was doing fine job itself, it decided to do only few records as sample. I was not able to enforce it to migrate everything no matter how I wanted. I’m not sure if that is problem with the mcp (not being able to read what data exist), or maybe some optimizations implemented to Cursor. Later seems more better explanation anyway.
Anyway, for context of this, I should maybe mention that I did fall back giving the old data in xml instead mcp reading. The agent was supposed to delete old data from database, but even that failed. Eventually I did do it with chatgpt generated sql scripts that read directly from the xml, but while it is good idea itself, the problem was that I wanted LLM AI to recognize some things (written by humans to invoice lines) from there, that cannot be automated with mere sql script.