I built an MCP server that allows me to query my local library docs and return the most relevant information for the query. The docs are in Markdown format.
I instruct Cursor (in the rules) to always run the MCP tool before replying and to take the returned data into consideration before coding.
The tool gets called but the data is not being taken into consideration at all. It just keeps on coding based on the outdated training data of the underlying model.
I’m unclear how Cursor uses the data returned or it if uses it at all? I see examples returning simple objects for weather data and the like, but how about a larger corpus of information to add to the context window?