Feedback Loop MCP

🚀 Feedback Loop MCP

A simple MCP server that enables a human-in-the-loop workflow in AI-assisted development tools such as Cursor. This server allows you to execute commands, view their outputs, and directly provide textual feedback to the AI. It is also compatible with Cline and Windsurf.

⚠️ Important Note

This project is inspired by interactive-feedback-mcp by Fábio Ferreira (@fabiomlferreira).

✨ Features

• Cross-platform: Compatible with macOS, Windows, and Linux.
• Interactive UI: A modern and responsive interface for collecting feedback.
• Settings persistence: Save and restore UI preferences for each project.
• MCP integration: Seamlessly integrates with MCP-compatible AI assistants.
• macOS overlay support: Native support for overlay windows on macOS.
• Dynamic Quick Feedback: Provide custom, clickable quick feedback options directly through the MCP tool call, enabling context-specific suggestions.

📷 Screenshot

The feedback collection interface with macOS vibrancy effects

📦 Installation

🚀 Quick Start with npx (Recommended)

The simplest way to use this MCP server is via npx:

npx feedback-loop-mcp

🌐 Global Installation

For frequent use, install it globally:

npm install -g feedback-loop-mcp
feedback-loop-mcp

👨‍💻 Local Development Setup

For development or customization:

1. Clone the repository:

   git clone <repository-url>
   cd feedback-loop-mcp
   

2. Install dependencies:

   npm install
   

3. Run in development mode:

   npm run dev
   

⚙️ MCP Server Configuration

📌 Cursor IDE

Add the following configuration to your Cursor settings (mcp.json):

{
  "mcpServers": {
    "feedback-loop-mcp": {
      "command": "npx",
      "args": ["feedback-loop-mcp"],
      "timeout": 600,
      "autoApprove": [
        "feedback_loop"
      ]
    }
  }
}

📌 Cline / Windsurf

Similar setup principles apply. Configure the server command in your MCP settings:

{
  "mcpServers": {
    "feedback-loop-mcp": {
      "command": "npx",
      "args": ["feedback-loop-mcp"]
    }
  }
}

📌 Claude Desktop

Add to your Claude Desktop configuration:

{
  "mcpServers": {
    "feedback-loop-mcp": {
      "command": "npx",
      "args": ["feedback-loop-mcp"]
    }
  }
}

💻 Usage

The feedback_loop tool accepts the following arguments:

• project_directory (string, required): The full path to the project directory.
• prompt (string, required): A combined summary and question that describes what has been done and requests specific feedback.
• quickFeedbackOptions (array of strings, optional): A list of predefined feedback strings that will be displayed as clickable options in the UI for faster feedback.

Example tool call from an AI assistant:

feedback_loop_mcp({
  project_directory: "/path/to/your/project",
  prompt: "I implemented the new user authentication flow with JWT. Does this approach meet your requirements?",
  quickFeedbackOptions: [
    "Yes, looks good!",
    "Needs minor changes to the error handling.",
    "Please try a different approach."
  ]
});

🏃 Running the Server

Via npx (Recommended)

npx feedback-loop-mcp

Via Global Installation

feedback-loop-mcp

Local Development

npm start

📋 Command Line Arguments

The application accepts the following command-line arguments:

• --project-directory <path>: Set the project directory.
• --prompt <text>: Set the initial prompt/summary text.

Example:

npm start -- --project-directory "/path/to/project" --prompt "Please review this code"

🛠️ Available Tools

The MCP server provides the following tool:

• feedback_loop: Displays a UI for collecting user feedback and returns the response.

Example usage in AI assistants:

{
  "tool_name": "feedback_loop",
  "arguments": {
    "project_directory": "/path/to/your/project",
    "summary": "I've implemented the changes you requested and refactored the main module."
  }
}

📝 Prompt Engineering

For optimal results, add the following to your custom prompt in your AI assistant:

Whenever you want to ask a question, always call the MCP feedback_loop tool.
Whenever you're about to complete a user request, call the MCP feedback_loop tool instead of simply ending the process.
Keep calling the feedback_loop tool until the user's feedback is empty, then end the request.

This ensures that your AI assistant uses this MCP server to request user feedback before marking tasks as completed.

🌟 Benefits

By guiding the assistant to check in with the user instead of making speculative, high-cost tool calls, this module can significantly reduce the number of premium requests (e.g., OpenAI tool invocations) on platforms like Cursor. In some cases, it helps consolidate up to 25 tool calls into a single, feedback-aware request, saving resources and improving performance.

Built applications will be available in the dist directory.

📁 Project Structure

feedback-loop-mcp/
├── main.js              # Main Electron process
├── preload.js           # Preload script for secure IPC
├── package.json         # Project configuration
├── README.md           # This file
├── assets/             # Static assets
│   └── feedback.png    # Application icon
├── renderer/           # Renderer process files
│   ├── index.html      # Main UI
│   ├── styles.css      # Styling
│   └── renderer.js     # UI logic
└── server/             # MCP server
    └── mcp-server.js   # Node.js MCP server

⚙️ Configuration

The application automatically saves settings using Electron's built-in storage:

• General settings: Window size, position, and UI preferences.
• Project-specific settings: Command history and project-specific configurations.

Settings are stored in the standard application data directory for each platform.

📋 Features Overview

📥 Feedback Collection

• Rich text feedback input.
• Automatic saving of feedback.
• JSON output format for easy integration.
• Timestamp and project information included.

📝 Prompt Engineering

Crafting effective prompts is crucial for obtaining useful feedback. When using the prompt parameter, aim to:

1. Be concise but complete: Describe what was done and what feedback you need in a single, clear statement.
2. Include context: Mention key implementation details or choices that affect the feedback you're seeking.
3. Ask a specific question: End with a clear question that guides the user to provide the feedback you need.
4. Frame the question properly: Instead of "Does this look good?" try "Does this implementation address your requirements for X?"

Example of an effective prompt:

I've implemented user authentication using JWT with a 24-hour expiration and refresh tokens. Does this approach provide the security level you need for this application?

🚀 Quick Feedback Options

When defining quickFeedbackOptions, consider:

• Provide a maximum of 3 - 5 options for the best user experience.
• Include both positive and constructive feedback options.
• Make options specific to the context, not generic.
• Order options from most positive to most critical.

Example of well-designed quick feedback options:

quickFeedbackOptions: [
  "Perfect, exactly what I needed",
  "Good approach, but please add more validation",
  "I'd prefer a different authentication method",
  "Let's discuss alternatives in more detail"
]

👨‍💻 Development

For development and build information, see DEVELOPMENT.md.

🐞 Troubleshooting

🛠️ Common Issues

1. MCP server not connecting: Ensure the server is running and the configuration is correct.
2. npx command not found: Make sure Node.js and npm are properly installed.
3. Permission errors: On Unix systems, you may need to make the binary executable.

🐛 Debug Mode

Run with debug output:

DEBUG=* npx feedback-loop-mcp

📄 License

MIT License - see package.json for details.