MCP Feedback Enhanced

MCP Feedback Enhanced is an MCP server that brings a human-in-the-loop workflow to your AI interactions. It makes the AI confirm things with you instead of just guessing and potentially going off the rails. This approach can significantly reduce unnecessary operations.

When your AI tool needs input or to verify a step, MCP Feedback Enhanced steps in. It figures out your environment, pops up an interface for you to give commands, text feedback, or even upload images. Then, it sends your feedback back to the AI so it can proceed, adjust, or stop based on what you said.

Features

• 🖥️ Dual Interface System – Qt GUI for local environments and Web UI for SSH remote setups with automatic environment detection
• 🎨 Modular Architecture – Completely refactored design with centralized management and modern themes
• 🖼️ Image Support – Upload PNG, JPG, JPEG, GIF, BMP, and WebP files via drag-and-drop or clipboard paste (Ctrl+V)
• 🌏 Multi-language Support – English, Traditional Chinese, and Simplified Chinese with smart system detection
• ⌨️ Keyboard Shortcuts – Ctrl+Enter to submit feedback, Ctrl+V for direct image pasting
• 🔄 Smart Resource Management – Automatic timeout handling and UI cleanup mechanisms
• 🎯 Cost Optimization – Reduce multiple AI tool calls into single feedback requests

Use Cases

• Code Review Workflows – Stop AI from making assumptions about code changes; get explicit user approval for modifications, refactoring decisions, or architectural choices before implementation
• File Management Operations – Prevent AI from automatically deleting, moving, or overwriting files by requiring user confirmation for any destructive file system operations
• Configuration Updates – Ensure AI doesn’t modify critical configuration files, database schemas, or deployment settings without explicit user review and approval
• Multi-step Development Tasks – Break complex development workflows into manageable chunks where users can provide feedback and course corrections at each stage

How to Use It

1. Install using uv package manager and run a quick functionality test:

# Install uv if not already installed
pip install uv
# Quick test to verify installation
uvx mcp-feedback-enhanced@latest test

2. Add the server configuration to your MCP-compatible client.

Basic Configuration (Recommended):

{
  "mcpServers": {
    "mcp-feedback-enhanced": {
      "command": "uvx",
      "args": ["mcp-feedback-enhanced@latest"],
      "timeout": 600,
      "autoApprove": ["interactive_feedback"]
    }
  }
}

Advanced Configuration (Custom Environment):

{
  "mcpServers": {
    "mcp-feedback-enhanced": {
      "command": "uvx",
      "args": ["mcp-feedback-enhanced@latest"],
      "timeout": 600,
      "env": {
        "FORCE_WEB": "true",
        "MCP_DEBUG": "false"
      },
      "autoApprove": ["interactive_feedback"]
    }
  }
}

3. Configure behavior using these environment variables:

• FORCE_WEB – Set to “true” to force Web UI usage (default: “false”)
• MCP_DEBUG – Enable debug mode with “true” (default: “false”)
• INCLUDE_BASE64_DETAIL – Include full Base64 encoding for images with “true” (default: “false”)

4. Configure your AI assistant with these rules for optimal human-in-the-loop workflow:

# MCP Interactive Feedback Rules
1. During any process, task, or conversation, whether asking, responding, or completing stage tasks, must call MCP mcp-feedback-enhanced.
2. When receiving user feedback, if feedback content is not empty, must call MCP mcp-feedback-enhanced again and adjust behavior based on feedback.
3. Only when user explicitly indicates "end" or "no more interaction needed" can you stop calling MCP mcp-feedback-enhanced, then the process is complete.
4. Unless receiving end command, all steps must repeatedly call MCP mcp-feedback-enhanced.
5. Before completing the task, use the MCP mcp-feedback-enhanced to ask the user for feedback.

FAQs

Q: How does this reduce AI platform costs?
A: Instead of AI making multiple speculative tool calls that might be wrong, the server consolidates these into single feedback requests. Users can approve or reject actions upfront, preventing costly iteration cycles and reducing total API calls by up to 96%.

Q: I’m seeing an “Unexpected token ‘D'” error. What’s up?
A: This usually means debug output is interfering. Try setting the environment variable MCP_DEBUG=false, or just remove it if you set it previously.

Q: Image uploads are failing.
A: Double-check the file size; it needs to be 1MB or less. Also, ensure it’s a supported format: PNG, JPG, JPEG, GIF, BMP, or WebP.

Q: The Web UI isn’t starting when I expect it to.
A: You can try forcing it by setting the environment variable FORCE_WEB=true. If it still doesn’t show, check your firewall settings to make sure it’s not blocking the connection.

Q: Does this work with Gemini Pro 2.5 for image analysis?
A: There’s a known compatibility issue with Gemini Pro 2.5 not correctly parsing uploaded image content. Claude-4-Sonnet handles image analysis much better, so we recommend using Claude models when you need image understanding capabilities.