Because the question travels as a tool call, several rounds of back-and-forth fit inside a single request — the author's argument for why asking beats guessing wrong and being asked to try again. It needs a desktop session: this is a Qt window on your machine, so it will not work on a headless host.
A desktop feedback window driven by one tool. When the assistant hits an ambiguity it calls `interactive_feedback`, a three-pane window opens on your machine, and your answer returns inside the same request.
- `interactive_feedback` — the assistant pauses mid-task and puts the question to you
- Predefined options you can choose with `Ctrl+1-5`, so a decision costs one keystroke
- A three-column window: message content, suggested options, and project information with Git status
- A configuration file at `~/.interactive_feedback_mcp/config.json` for theme, language, font, window size and pane ratios
- Separate log files for the main log, errors, performance and project context, with a `manage_logs.py` script to view, search, analyse and export them
Python 3.10+ with PySide6 and FastMCP, launched through uv: `uv --directory /absolute/path/to/interactive-feedback-mcp run server.py`. Give the entry a generous timeout — the sample config uses 600 — and add `interactive_feedback` to autoApprove. Windows 10+, macOS 12+ or Ubuntu 20.04+.