I use Cursor as a desktop editor. The files I want an agent to change are already on this computer, in the folder I have open.
Agents launched from that desktop app often run as Cloud Agents. A Cloud Agent is a remote machine. It cannot read or write the folder on my computer. The chat then says the work is “in the cloud” and that it cannot put the file on my machine. From the desktop window, that is indistinguishable from the local agent until the task has already failed.
That split is useful when I ask for it. A cloud machine is a reasonable place to run a job against a GitHub repo, or to keep working after the laptop is closed. It is a bad default for an agent I started because I am sitting in a local project.
What I expected:
- An agent started from the desktop app, with a folder open, reads and writes that folder.
- A cloud run is a separate action, labeled before the first prompt, with the limit stated in the product: this run cannot see your disk.
- If a chat is already on the cloud machine, the window shows that, instead of leaving the model to discover it and refuse the task.
- There is a direct action, “Run this on my computer,” that continues the same task against the open folder.
The October 6 changelog describes Projects as running on their own computer in the cloud, with a local agent only when the coordinator decides a test must happen on the machine. For local files that are not a hosted repo, that order is backwards. The computer in front of me is the one that has the work.
I am not asking to remove Cloud Agents. I am asking that the desktop app’s ordinary agent be local, and that cloud stay an explicit choice.