I recently noticed something unusual while working on a frontend component. I had my Figma design open in Brave Browser and Cursor IDE running at the same time. Cursor’s styling suggestions were surprisingly accurate and closely matched the design in Figma.
Initially, I assumed Cursor was simply inferring the design based on the surrounding code and project context. However, today I encountered a different situation while building a completely new component. The component contained text that had never been used anywhere in my project before. As I was manually typing that text, Cursor suggested the exact wording that appeared in the Figma design even though I had not copied or pasted it from Figma.
Has anyone else experienced something similar? Does Cursor have any documented integration or capability that allows it to access content from open browser tabs or design tools such as Figma or screen content? Is this behavior covered in Cursor’s terms, privacy policy, or documentation?
I’m curious if there’s a technical explanation for this or if others have observed the same behavior.
Sorry. No one’s talking about it, but I’ve watched my agent try every which way to access files that were supposed to be off limits - using system tools outside of the IDE. ex: the .env file
I wouldn’t be surprised if their agents (opus esp) are just straight reading everything everywhere in every other app/browser window.
Cursor isn’t reading your browser tabs or watching your screen. There’s no screen-capture or “look at whatever’s focused” behavior.
If you’ve got a Figma integration set up (Cursor can connect to design tools through MCP, Figma has one), then design context can flow in on purpose, through a connection you’ve authorized. Worth a look in your MCP/integrations settings.
Beyond that, Cursor already has a lot of context from your project: open files, recent edits, and the indexed codebase. Design systems repeat themselves a lot, so token names, spacing, and even copy often already live somewhere in your code or history, and the model is good at picking up on that.
The “exact wording” thing usually comes down to one of those: either a connected design source, or the text existing somewhere in your workspace you’ve forgotten about. Models are also just weirdly good at guessing the obvious label for a button or heading.
