MCP Feedback App

Interactive feedback UI that renders directly inside MCP hosts (Claude, Cursor, etc.) using the MCP Apps extension. Includes a macOS floating companion app for bidirectional feedback and pending message management.

When an AI agent calls the interactive_feedback tool, a feedback panel appears inline in the conversation — and optionally in the floating app. The user can respond from either side.

Features

MCP App Panel (in-conversation UI)

• Inline feedback UI — summary with Markdown rendering + input area, directly in the conversation
• Dark mode — auto-detects host theme (transparent background adapts to light/dark)
• Image support — paste (Cmd+V), drag & drop, file picker with preview
• Quick replies — Continue, LGTM, Fix, Pause one-click buttons
• Auto-append — automatically appends a configurable reminder to every feedback
• Keyboard shortcut — Ctrl/Cmd+Enter to submit feedback
• Blocking tool pattern — interactive_feedback blocks until user submits via the UI

Floating Companion App (macOS)

• Bidirectional feedback — both the MCP App panel and floating app receive AI summaries; respond from either side
• Pending messages — queue messages that redirect the agent mid-task (delivered via preToolUse hook)
• Per-session tabs — manage multiple agent sessions with project directory info
• Chat history — full conversation history loaded from events.jsonl per session
• Image support — paste images (Cmd+V), file picker, thumbnails in chat, click to preview
• Global hotkey — Cmd+Shift+M to toggle the panel
• Floating button — collapsible/expandable, draggable, always-on-top
• State persistence — window position, size, expanded state, input draft all persist across restarts
• Enter to send — Enter sends, Shift+Enter for newline, IME-safe
• Menu bar icon — quick access via status bar

Cursor Integration

• Auto-install — MCP server config, Cursor rules, and hooks are all auto-configured on first startup
• Subagent protection — hooks use per-conversation reference counting (subagentStart/subagentStop) to block feedback calls when subagents are active
• Session lifecycle — automatic session tracking via hooks (sessionStart, preToolUse, etc.)
• Event logging — all interactions logged to events.jsonl per session for chat history

Architecture

┌─────────────────────────────────────────────────┐
│          MCP Host (Cursor / Claude)             │
│  ┌──────────┐    ┌────────────────────┐         │
│  │ AI Agent │    │ MCP App Panel      │         │
│  │          │    │ (sandboxed iframe)  │         │
│  │ calls    │    │ sessionId embedded │         │
│  │ tool ────┼────┼─► renders inline   │         │
│  └──────────┘    └────────┬───────────┘         │
└────────────────────────────┼────────────────────┘
                             │ callServerTool
                             ▼
               ┌───────────────────────────┐
               │       MCP Server          │
               │  interactive_feedback     │ ← polls session file
               │  submit_feedback          │ ← writes resolved file
               │  get_system_info          │
               └───────────┬───────────────┘
                           │
               ┌───────────▼───────────────┐
               │  ~/.cursor-mcp-feedback/     │
               │    sessions/{uuid}.json   │  ← bidirectional channel
               │    sessions/{convId}/     │
               │      events.jsonl         │  ← chat history
               │      pending.json         │  ← queued messages
               │      meta.json            │  ← session metadata
               │    active-sessions.json   │
               └───────────┬───────────────┘
                           │ file watch
               ┌───────────▼───────────────┐
               │  Floating App (macOS)     │
               │  SwiftUI + AppKit         │
               │  - bidirectional feedback │
               │  - pending messages       │
               │  - per-session chat UI    │
               └───────────────────────────┘

Installation

Quick Install (npm)

npm install -g cursor-mcp-feedback

On install, postinstall automatically configures Cursor (mcp.json, rules, hooks). Just restart Cursor and the MCP server is ready.

From Source

git clone [email protected]:cash/incubating/cusor-feedback-v2.git cursor-mcp-feedback
cd cursor-mcp-feedback
npm install
npm run build

Floating Companion App (macOS)

cd floating-app

# Option A: Build DMG for distribution
./build-dmg.sh        # Creates CursorMCPFeedback.dmg (~800KB)
# DMG contains: .app + Install.command (handles xattr) + /Applications shortcut

# Option B: Build .app directly
./build-app.sh        # Creates CursorMCPFeedback.app
cp -r CursorMCPFeedback.app /Applications/

To auto-launch on login, add CursorMCPFeedback to System Settings > General > Login Items.

CLI: Queue Pending Messages

# List active sessions
cursor-mcp-feedback queue sessions

# Queue a message to the most recent session
cursor-mcp-feedback queue "please focus on the API layer"

# Queue to a specific session
cursor-mcp-feedback queue "fix the bug" --session <session-id>

# List pending messages
cursor-mcp-feedback queue list

# Clear pending messages
cursor-mcp-feedback queue clear

Configure your MCP host

Cursor (auto-configured)

On first startup (or npm install -g), the MCP server automatically configures itself in ~/.cursor/mcp.json with the correct absolute paths. It also auto-installs:

• Cursor rule (~/.cursor/rules/cursor-mcp-feedback.mdc) — instructs the agent to call interactive_feedback
• Cursor hooks (~/.cursor/hooks/) — subagent protection + pending message delivery + event logging
• hooks.json entries — registers subagentStart, subagentStop, beforeMCPExecution, afterMCPExecution, preToolUse hooks

All auto-installed files are kept in sync on every startup (hash-based diffing, idempotent).

To manually configure instead, add to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "cursor-mcp-feedback": {
      "command": "node",
      "args": ["/absolute/path/to/cursor-mcp-feedback/dist/main.js"],
      "timeout": 86400,
      "env": {
        "MCP_FEEDBACK_TIMEOUT": "86400"
      }
    }
  }
}

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "cursor-mcp-feedback": {
      "command": "node",
      "args": ["/absolute/path/to/cursor-mcp-feedback/dist/main.js"],
      "timeout": 86400,
      "env": {
        "MCP_FEEDBACK_TIMEOUT": "86400"
      }
    }
  }
}

Cursor Hooks

| Hook | Event | Purpose | |------|-------|---------| | block-cursor-mcp-feedback.js | subagentStart | Increment active subagent counter for parent conversation | | block-cursor-mcp-feedback.js | subagentStop | Decrement active subagent counter | | block-cursor-mcp-feedback.js | beforeMCPExecution | Deny cursor-mcp-feedback MCP calls when subagents are active | | block-cursor-mcp-feedback.js | afterMCPExecution | Log feedback_request/response events to events.jsonl | | consume-pending.js | preToolUse | Consume pending messages and inject as agent feedback |

Keyboard Shortcuts

| Shortcut | Context | Action | |----------|---------|--------| | ⌘+Enter / Ctrl+Enter | MCP App Panel | Submit feedback | | Enter | Floating App | Send message / feedback | | Shift+Enter | Floating App | Insert newline | | ⌘+Shift+M | Global | Toggle floating panel | | Esc | Floating App | Collapse to floating button | | ⌘+V | Floating App | Paste image from clipboard |

UI Settings

| Setting | Default | Persistence | Description | |---------|---------|-------------|-------------| | Auto-append reminder | On | ~/.cursor-mcp-feedback-settings.json | Appends text to every feedback | | Auto-append text | After completing, call interactive_feedback... | Server-side | The text appended after each feedback |

Server Configuration

| Environment Variable | Default | Description | |---------------------|---------|-------------| | MCP_FEEDBACK_TIMEOUT | 86400 (24h) | Server-side timeout in seconds |

Development

npm run build:ui     # Build the UI (Vite → single-file HTML)
npm run build:server # Compile TypeScript server
npm run build        # Both
npm run dev          # Run with tsx (dev mode, stdio)

Project Structure

src/
  App.tsx              # Main React component (MCP App Panel UI)
  mcp-app.tsx          # React entry point
  styles/app.css       # Styles (light + dark theme)
server.ts              # MCP server (tools + resources)
main.ts                # Entry point (stdio/HTTP + auto-install)
session-store.ts       # File-based session CRUD
logger.ts              # File logging
hooks/
  block-cursor-mcp-feedback.js  # Subagent protection + event logging
  consume-pending.js         # Pending message delivery
  session-utils.js           # Shared session utilities
floating-app/
  Sources/
    FloatingApp.swift        # App delegate, menu bar, window management
    FloatingPanel.swift      # NSPanel subclass (keyboard, IME handling)
    FloatingButtonView.swift # Collapsed floating button
    PendingPanelView.swift   # Main panel UI (chat, input, tabs)
    SessionManager.swift     # Session data, chat history, file I/O
    FileWatcher.swift        # File system event monitoring
    ImagePreviewWindow.swift # Image preview via system Preview.app
  Package.swift              # Swift Package Manager config
rules/
  cursor-mcp-feedback.mdc      # Cursor rule (auto-installed)

MCP Tools

| Tool | Visibility | Description | |------|-----------|-------------| | interactive_feedback | model | Presents summary (Markdown), blocks until user responds | | submit_feedback | app-only | Submits user feedback with optional images (called from UI) | | get_system_info | model | Returns OS, architecture, Node.js version |

Prerequisites

• Node.js >= 18
• macOS 13+ (for floating companion app)
• MCP host with MCP Apps support

License

MIT