The economics are the argument: an assistant that is unsure will branch and try things, each attempt a paid request, and asking instead collapses that into one exchange. The panel earns the extra step by carrying more than text — pasted screenshots go back as images the model can read, and command output comes back in the same reply.
A human-in-the-loop server for Cursor and similar tools. When the assistant would otherwise guess, it opens a Qt window showing what it has done, waits for your reply, and hands both your feedback and any command output back. The README is in Chinese, with an English version alongside it.
- `interactive_feedback` takes a `project_directory` and a `summary`, and returns `logs` plus your typed `interactive_feedback`
- Quick-reply buttons for the common answers — approve, ask for a small change, work through the remaining checklist
- RIPER-5 buttons that push a structured pass: research, innovate, plan, execute, review
- Images by drag-and-drop or Ctrl+V, compressed to a target you set and sent as Base64 for the model to look at
- A command panel with live output, process start and stop, and per-project command history
- Settings stored per project — the run command, button layout, language and image options are remembered separately for each directory
- An interface that switches between Chinese and English on the spot
Python 3.11 or newer and `uv`, plus a desktop session — this opens a real window, so it is not something to run on a headless host. Clone it, `uv venv`, `uv sync`, then point Cursor's `mcp.json` at `server.py` through `uv`. Two config details matter: raise the client timeout, since the call waits for a person, and auto-approve `interactive_feedback` so the prompt itself does not need approving. Version 1.0.0.