Maybe you have too many MCP servers enabled. Try disabling the ones you don’t need at the moment. I also hope we’ll release an update soon where you can disable individual MCP server tools, which I think will help in the future.
Just got the issue using figma mcp. Just a simple prompt ask it to list small part of metadata. Looks like the response of mcp tool is also taken into account of conversation weight. The new conversation lost the previous context.
same here, this is unusable, my team and I will migrate to claude code for now, the cursor left us with no choice
Literally installed very first MCP, same error on brand new conversation (Reddit MCP - post search test). Not too many MCP, this is a single MCP install on someone that’s never run one before. Mode: Agent Model: Auto, fail. Model Claude-3.5-sonnet, fail. Model: claude-3.7-sonnet, pass. Model: deepseek-v3, pass. All other models fail to load up MCP for me.
Looks like for MCP calls, the Auto needs set to the only free model that can use them, deepseek-v3, without the “Your conversation is too long” error.
I’ll use this as a work around for now, but clearly the auto just needs fixed to move MCP calls to v3, at a minimum. However, the Cursor Models page does not list this model as an agent, which is confusing, as I thought only an Agent could use MCP.
Thanks for sharing the workaround. From my side,it at least works around 4-5 prompts then only the error appears
I’m getting this everytime the agent calls supabase’s mcp list_tables function (I have a lot of tables). It makes the agent unusable. It should find the key components of the output of each step and summarize it for the main agent’s line of thinking, similar to how Roocode does it.
I guess the only solution is to use one chat per feature, after it is implemented, you delete that chat and start another chat for a different feature. Currently this is what am doing, i have not gotten the error since