Hey, thanks for the report. What you’re seeing is the raw clipboard contents, meaning the image bytes shown as text, getting pasted into the terminal input. On Windows, the CLI doesn’t have a handler for clipboard image paste yet, so when you press Ctrl+V with an image in the clipboard, its binary content gets dumped into the input. On macOS and Linux this already works correctly. Windows is next, but I can’t share an ETA yet.
About what gets sent out:
- Cursor doesn’t upload anything in the background. The server only gets what you explicitly send as a prompt by pressing Enter.
- If you didn’t send the message, no tokens were used and nothing was sent anywhere. You can just clear the input with
Ctrl+Cor restart the session. - If you did send it, then the model received the same garbage text you see in the input, but it’s still just text, not a separate data channel.
Workarounds for now:
- Save the image as a file and pass the path in the prompt, for example
analyze ./screenshot.png - Or use image paste in Cursor IDE itself or Cursor Agent in the browser. Pasting an image with
Ctrl+Vworks there.
Related threads for context:
- Pasting images not work on Window cursor cli
- Title: I hope Cursor CLI can support pasting images via Ctrl+V
If you see “web stream” even without pasting an image, meaning it appears right after running cursor agent with no clipboard actions, let me know your CLI version from cursor --version and the repro steps, and we’ll look into that separately.