The pattern this enables is a client rule that says: before you finish, ask. The assistant reports, you answer in the popup — often with a screenshot pasted straight from the clipboard — and it continues from your correction rather than its own assumption. The window has to appear on a screen you are looking at, so this is a desktop tool and nothing else.
A Python server that opens a dialog on your machine. The assistant reports what it just did, you reply with text, images, or both, and the reply goes back into the conversation — which lets a session continue instead of ending on the model's guess.
- A feedback dialog showing the assistant's report and collecting text plus several images — `collect_feedback`
- A quick single-image picker for when only a screenshot is needed — `pick_image`
- Format, dimensions and file size of an image — `get_image_info`
- Images pasted straight from the clipboard with Ctrl+V or chosen from disk, in PNG, JPG, JPEG, GIF, BMP or WebP
- Keyboard handling built for speed: Enter submits, Esc cancels
Python 3.10+ and a desktop session — this opens a real window, so it will not work on a headless machine. Install uvx, then have the client launch `uvx` with the `mcp-feedback-collector` package. `MCP_DIALOG_TIMEOUT` sets how long the dialog waits in seconds and defaults to 300; the project suggests 600 for normal work and 1200 for long operations. Set `PYTHONIOENCODING` to `utf-8`. The package is `mcp-feedback-collector` 2.0.0.
One command — uvx mcp-feedback-collector