beek
July 11, 2025, 3:34pm
1
Describe the Bug
When I provide a task for the cursor agent, any attempt to use any basic tool (reading files, writing files, command line, etc…) returns an error message to the agent. There is no detail in the error message which might explain what is happening. I’ve already checked for permissions, its all good. I’ve restarted cursor, and my computer. Still the same issue.
Steps to Reproduce
Ask cursor agent to do any task. Ask it to explain the error it is receiving.
Expected Behavior
I expect cursor to be able to use its own basicl tools.
Screenshots / Screen Recordings
Operating System
MacOS
Current Cursor Version (Menu → About Cursor → Copy)
Version: 1.2.4 (Universal)
VSCode Version: 1.99.3
Commit: a8e95743c5268be73767c46944a71f4465d05c90
Date: 2025-07-10T16:55:16.443Z
Electron: 34.5.1
Chromium: 132.0.6834.210
Node.js: 20.19.0
V8: 13.2.152.41-electron.0
OS: Darwin arm64 24.5.0
Does this stop you from using Cursor
Yes - Cursor is unusable
1 Like
Are you sure that the chat is in Agent
mode and not Ask
?
beek
July 11, 2025, 3:36pm
3
Yes, chat is in agent mode.
Hey, which model are you using? Also, is this happening in a new chat?
beek
July 11, 2025, 5:02pm
5
I am using gemini-2.5-pro max.
It happens in a new chat or an old one. I’ve tried in multiple different projects also.
I just changed the model to claude-4-sonnet and that model is working. Changing back to gemini makes the tools not work again.
2 Likes
Also getting the same. It seems this is a Gemini issue though - working with Opus it was able to read files for me.
Not a single frontier model can use the todo list feature for me. Apparently it’s not available in their scope. How can we test tool availability?
beek
July 14, 2025, 1:56pm
8
The gemini-2.5-pro model STILL cannot use any cursor tools.
What is going on here?
1 Like
R_K_FB
July 14, 2025, 4:58pm
9
The gemini-2.5-pro model STILL cannot use any cursor tools.
Confirm. Gemini-2.5-pro-preview-06-05
and Gemini-2.5-flash
are working though.
I’d be surprised if this wasnt a censoring_on_crack style bug.
A Claude example is that I have a copy of an unredacted little black book that I wanted to do interesting data sciencey things with…
But Claude recognizes what it is - even in partial PDF chunks and REFUSES to work with the data.
It simply just cancels the request and acts like you didnt even type it in…
This is the scary Dark_Censorship we have no control over.
Gemini refuses to do certain deep dives on certain people…
I wouldnt be surprise if this is just overly Nerf’d behavior on Gemini’s part.