I found LLM's make a bucket load of mistakes, so I made my own mcp server to try fix these problems. I included tool stats and persistent project notes, to steer tool use and optimization, apparently it is not a thing, and all this usefull stuff just goes out the context windows and is lost.
I created this document on some of the problems with Cline and Claude code etc.
I used stats to design my tools to limit token use and drop failed edits quite a bit.
might be worth a read.
There is also sample working code for examples of these ideas.
I found LLM's make a bucket load of mistakes, so I made my own mcp server to try fix these problems. I included tool stats and persistent project notes, to steer tool use and optimization, apparently it is not a thing, and all this usefull stuff just goes out the context windows and is lost.
I created this document on some of the problems with Cline and Claude code etc.
I used stats to design my tools to limit token use and drop failed edits quite a bit. might be worth a read. There is also sample working code for examples of these ideas.
https://github.com/professor-jonny/pulsar-edit-mcp-server/bl...