I’m not sure if this is a Claude bug or Cursor bug, but I never had it happen to me in Claude web interface so I assume it’s cursor. It is unrelated to the MCP tool, it happens without custom MCP tools as well. I didn’t notice it happen until the last few days, now it just randomly happens a few times per day - possibly related to a new Cursor version I updated to. Unfortunately don’t know how to reproduce it, it just sometimes happens and sometimes doesn’t.
The issue is (see screenshot) that the exact same chain of thought is repeated multiple times. The model seems to behave fine in it’s normal text response, but the chain of thought is repeated word-to-word identical, sometimes 5 times or more during the request. It never returns to a different text in the chain of thought until I ask a new question, then it works fine again.
It’s not a serious bug, I can still use Cursor just fine. The model performance doesn’t seem any worse with this bug.
Cursor Version: 1.0.0
MacOS Version: 15.4 (Apple Silicon)
5 Likes
Hey, this is likely a consequence of the MCP server not returning an output in the format Cursor / the AI expects, so it continues to re-think over the issue as it doesn’t think anything has changed.
Does the same happen if you don’t use the MCP server and attempt the same task?
Honestly I don’t remember, next time it happens I will check if it was using the MCP or not and I will get back to you if it ever happens without the MCP too. Thanks for the help!
I was in Agent mode, no MCP was used in the conversation, although I still have the MCP server enabled so the Agent still knows about the MCP even if it didn’t use it. I doubt it’s the MCP though. It repeated the same chain of thought 11 times in one response, then it switched to another chain of thought which it repeated 5 times, still in that same response. I provided some screenshots, the rest of the conversation is similar, it works as expected and does what I asked it to do, but the chain of thought repeats (the shorter one 11 times and the longer one 5 times).
1 Like
Getting this issue as well, without MCP tools. It’s very frustrating, it’s in a loop, retaking the actions it just did…
I am having the same issue. Started last night.
Looks like its only happening on claude sonnet 4 at the moment. Gemini 2.5 Pro is not showing the same issue. going to investigate more.
I also noticed that situation (using Claude 4 sonnet) since I updated Cursor to 1.0.1 or later. I wished to upgrade the version to fix it, but it didn’t help (until now it’s 1.1.3). So depressed 
I’m constantly having this issue with claude-4-sonnet, the thinking tokens displayed are actually IDENTICAL, to the point that it makes me believe it might be a Cursor bug rather than model behavior. It happens with and without MCPs, so I don’t think that’s related.
This seems to only happen with claude-4-sonnet Thinking without MAX mode.
I can also confirm this is happening with Claude-4-sonnet Thinking on the latest version. MCP is not being used. The worst part is it appears to be progressing in the problem, but the old thinking keeps polluting the changes, so it doesn’t learn and react correctly.
Me too. Claude-4-sonnet keeps repeating verbatim it’s initial thinking even after we have moved on several steps. Strangely it keeps making progress despite this weird loop.
