cursor-better-feedback

English | 中文

MCP feedback tool with interactive UI for Cursor. Replaces the built-in AskQuestion with a more robust feedback mechanism using MCP Apps.

| Waiting for feedback | After submission | |:---:|:---:| | | |

Features

• Interactive feedback UI rendered directly in the Cursor chat panel (via MCP Apps iframe)
• Image paste: Ctrl+V to attach screenshots (PNG/JPEG/WebP/GIF, up to 5 images, 5MB each)
• Markdown rendering in message area (raw HTML stripped for security)
• Configurable timeout and font size via environment variables
• Host theme/style adaptation
• Keyboard shortcut: Ctrl+Enter / Cmd+Enter to submit
• Dual transport: stdio (default, for Cursor) + HTTP (development only)

Installation

Add to your Cursor MCP configuration (.cursor/mcp.json):

{
  "mcpServers": {
    "feedback": {
      "command": "npx",
      "args": ["cursor-better-feedback"],
      "env": {
        "FEEDBACK_TIMEOUT": "1200",
        "FEEDBACK_FONT_SIZE": "12px"
      }
    }
  }
}

Or use a local path:

{
  "mcpServers": {
    "feedback": {
      "command": "node",
      "args": ["/path/to/cursor-better-feedback/dist/main.js"]
    }
  }
}

Configuration

| Variable | Default | Description | |----------|---------|-------------| | FEEDBACK_TIMEOUT | 1200 | Default timeout in seconds (60-3600) | | FEEDBACK_FONT_SIZE | 12px | UI font size (e.g. 12px, 0.875rem) |

How It Works

1. The LLM calls feedback(message="...") when it needs user input
2. Cursor renders the feedback UI in an iframe (MCP Apps)
3. User types feedback, optionally pastes images with Ctrl+V
4. User clicks Submit (or presses Ctrl+Enter)
5. The UI calls submit_feedback to resolve the pending tool call
6. The LLM receives the feedback text (and any attached images) and continues

Recommended Cursor Rule

Add this rule to .cursor/rules/feedback.mdc to enforce feedback-driven interaction with automatic retry:

---
description: "Feedback MCP interaction protocol with resilience"
globs:
alwaysApply: true
---

# Feedback Interaction Protocol

## Interaction Rules (Mandatory)

- **MUST** use `feedback` tool for confirmations/choices — never use plain text questions
- During discussion: natural language only, no large code blocks (pseudocode/minimal snippets OK)
- When information is unclear or missing: ask the user immediately, never assume

## End of Turn (Mandatory)

After every response, **MUST** call `feedback`: "还有其他需要吗？你也可以直接输入下一个指令。"
- If user says "end" / "done" / "no more" → end conversation
- Otherwise → treat response as next instruction
- **Only exception**: user already explicitly ended this turn

## Resilience

1. Always call `feedback` tool first
2. If `Tool not found` or connection error:
   - Sleep 5 seconds, retry up to **3 times**
3. All retries failed → fall back to built-in `AskQuestion`

Network fluctuations may cause Cursor to temporarily disconnect MCP servers.
The feedback server (stdio) itself is unaffected; Cursor auto-reconnects after recovery.

Limitations

• Only one pending feedback session at a time. A new feedback call cancels any previous pending session. Multi-agent concurrent feedback is not supported.
• The upper/lower padding in the UI is controlled by Cursor's iframe container and cannot be adjusted from within the app.

Requirements

• Cursor v2.6+ (MCP Apps support required)
• Node.js >= 18

Transport Modes

• stdio (default): Used for Cursor integration. The server communicates via stdin/stdout.
• HTTP (--http flag): Development/testing only. Binds to 127.0.0.1:3001. Not intended for production use.

Local Development

npm install
npm run build
npm run serve          # stdio mode (default)
npm run serve:http     # HTTP mode (localhost only, for testing)
npm run dev            # Development mode (watch + HTTP)

Testing with basic-host

# Terminal 1: Start the server in HTTP mode
npm run build && npm run serve:http

# Terminal 2: Start the basic-host test harness
cd /tmp/mcp-ext-apps/examples/basic-host
npm install
SERVERS='["http://localhost:3001/mcp"]' npm start

# Open http://localhost:8080 in your browser

License

MIT