Minidoracat/mcp-feedback-enhanced

What is MCP Feedback Enhanced?

MCP Feedback Enhanced is a server‑side tool that plugs into the Model Context Protocol (MCP) ecosystem. It adds a human‑in‑the‑loop feedback UI for long‑running AI tasks. When an AI model (e.g., Cursor, Cline, Windsurf, Augment, Trae) needs confirmation or clarification, the model calls the mcp-feedback-enhanced command. The tool then launches a Web UI (or an optional Tauri‑based desktop app) where the user can review the request, type a response, upload images, or let a timer auto‑submit. The response is sent back to the AI over a WebSocket, letting the model continue with the new information.

The project is a fork of the original interactive-feedback-mcp implementation, maintained by Minidoracat. It focuses on security (removing unsafe command execution, tightening WebSocket origin checks) and on making the UI work reliably in SSH‑remote, WSL, and local environments.

---

Core ideas

Idea	Why it matters
Feedback‑oriented workflow	Inserts explicit human checkpoints, reducing drift and re‑work in long AI‑driven processes.
Dual interface	A lightweight browser UI works everywhere; a Tauri desktop wrapper provides a native‑look app for Windows/macOS/Linux (maintenance‑only, slated for removal in v3).
Environment awareness	Detects SSH remote sessions and WSL automatically, choosing the appropriate UI mode without extra configuration.
Session tracking	Saves each feedback session locally (JSON/CSV/Markdown) and shows real‑time statistics, making it easy to audit or share results.
Security‑first updates	v2.6.1 removes a dangerous command‑execution feature and validates WebSocket Origin headers, closing two high‑severity vulnerabilities.

---

Who would use it?

• Developers building AI‑assisted tools that need occasional user confirmation (e.g., code‑generation assistants, data‑labeling pipelines).
• Teams using MCP‑compatible clients (Cursor, Cline, etc.) who want a consistent UI for feedback across local, remote, or containerised environments.
• Researchers prototyping interactive LLM workflows who need a quick way to capture and replay human input.

---

Main features (as of v2.6.0)

• Web UI – runs on http://127.0.0.1:8765 by default, no GUI dependencies, works over SSH port‑forwarding.
• Desktop app – Tauri wrapper around the same UI (maintenance‑only, will be removed in v3).
• Prompt management – create, edit, delete reusable prompts; view usage stats; automatic sorting.
• Auto‑timed submit – configurable timer (1 s – 86400 s) with pause/resume/cancel controls.
• Session history – local file storage, export to JSON/CSV/Markdown, live statistics.
• WebSocket connection monitoring – shows status, auto‑reconnects, quality indicators.
• Rich Markdown display – AI summaries rendered with headings, code blocks, lists, etc.
• Audio & system notifications – optional sound effects and OS‑level alerts for events like auto‑commit or timeout.
• Internationalisation – UI available in English, Traditional Chinese, Simplified Chinese; language can be forced via MCP_LANGUAGE.
• Image support – drag‑and‑drop or clipboard paste of PNG/JPG/GIF/BMP/WebP of any size.

---

Getting started (quick‑run)

# Install the uv runner (if you don’t have it)
pip install uv

# Run the latest version (v2.6.1 at time of writing)
uvx mcp-feedback-enhanced@latest

The command starts the MCP server, launches the Web UI, and prints the URL (e.g., http://127.0.0.1:8765). Point your browser there, or set "MCP_DESKTOP_MODE": "true" in your MCP config to open the native desktop window.

---

Configuration snapshot

{
  "mcpServers": {
    "mcp-feedback-enhanced": {
      "command": "uvx",
      "args": ["mcp-feedback-enhanced@latest"],
      "timeout": 600,
      "env": {
        "MCP_DEBUG": "false",
        "MCP_WEB_HOST": "127.0.0.1",
        "MCP_WEB_PORT": "8765",
        "MCP_LANGUAGE": "en"
      },
      "autoApprove": ["interactive_feedback"]
    }
  }
}

*Replace uvx mcp-feedback-enhanced@latest with a pinned version (e.g., @2.8.0) if you want to keep the desktop app after v2.8.

---

Security & maintenance notes

• v2.6.1 removes the old “auto‑command” feature that let unauthenticated WebSocket clients run arbitrary shell commands. The server now only serves the UI and WebSocket; no command execution is possible.
• WebSocket origin validation prevents other sites from hijacking the connection.
• The desktop binary is large (~80 % of the package) and will be deprecated in v3; the Web UI is the recommended long‑term interface.
• The original “quota‑saving” claim (consolidating multiple round‑trips) is no longer valid after Cursor switched to token‑based pricing in June 2025. The tool’s value now lies in human checkpoints, not cost reduction.

---

Where to learn more

• Full changelog – RELEASE_NOTES/CHANGELOG.en.md
• Security details – SECURITY.md
• SSH remote usage guide – docs/en/ssh-remote/browser-launch-issues.md
• Example config files – examples/mcp-config-web.json and examples/mcp-config-desktop.json

---

TL;DR

MCP Feedback Enhanced is an MCP‑compatible server that shows a web‑based (or optional desktop) UI for humans to give quick feedback to AI models during long tasks. It works in local, SSH‑remote, and WSL environments, tracks sessions, supports images and markdown, and has recently been hardened against command‑execution and WebSocket hijacking vulnerabilities. Use it when you need a reliable, language‑aware human‑in‑the‑loop checkpoint for any MCP‑enabled AI client.