The pitch is concrete: by asking rather than branching into guesses, a stretch that would have taken up to 25 tool calls collapses into a single feedback-aware request — which on a metered platform is money as well as time.
A single-purpose server that opens a desktop window for human-in-the-loop review. The assistant calls it instead of guessing or instead of declaring a task finished, and waits for what you type.
- One tool, `interactive_feedback`, taking a `project_directory` and a `summary` of what the assistant just did, and returning your typed reply
- A command box in the same window: run something, see its output, and let that inform the feedback you send back
- Per-project settings kept in Qt's `QSettings` — the command to run, whether to execute it automatically on the next start for that project, and whether the command section is shown
- A "Save Configuration" button for the current command and the auto-run checkbox; the command section's visibility saves the moment you toggle it
Python 3.11 or newer and `uv`. Clone the repository, run `uv sync`, then start it with `uv run server.py`. In the client config, point `command` at `uv` with `--directory` set to your clone; the documented example sets a `timeout` of 600 and auto-approves `interactive_feedback`. Settings are stored in the platform's usual place — the registry on Windows, plist files on macOS, `~/.config` or `~/.local/share` on Linux — under organization `FabioFerreira` and application `InteractiveFeedbackMCP`. Compatible with Cursor, Cline and Windsurf.
One command — uvx interactive-feedback-mcp