⚡ Cursor Snap MCP

Snap to interrupt. Snap to screenshot.

One-click agent control for Cursor IDE.

Enhanced fork of Interactive Feedback MCP, combined with screenshot capture from Capture MCP, plus the first-ever Agent Interrupt mechanism.

🎯 Caught the agent going wrong? One click to interrupt and redirect — without burning a single extra request.

Installation

For Humans

Copy and paste this prompt to your Cursor Agent:

Install and configure cursor-snap-mcp by following the instructions here:
https://raw.githubusercontent.com/zrn6640/cursor-snap-mcp/main/docs/install-guide

Or read the Installation Guide, but seriously, let an agent do it.

For Cursor Agent

Fetch the installation guide and follow it:

curl -s https://raw.githubusercontent.com/zrn6640/cursor-snap-mcp/main/docs/install-guide

Manual (One-liner)

git clone https://github.com/zrn6640/cursor-snap-mcp.git && cd cursor-snap-mcp && uv sync && ./install.sh /path/to/your/project

Then restart Cursor.

Verify Installation

After installation, check Cursor Settings:

Hooks (Settings → Hooks):

✓ Configured Hooks (3)

preToolUse
  .cursor/hooks/interrupt-check.sh

subagentStart
  .cursor/hooks/interrupt-check-subagent.sh

afterMCPExecution
  .cursor/hooks/clear-interrupt.sh

MCP Tools (Settings → MCP):

✓ interactive-feedback
  Tools: interactive_feedback

Three Core Features

1. ⚡ One-Click Agent Interrupt (Unique)

Agent heading the wrong way? Don't wait for it to finish:

Click tray icon → Agent's next tool call blocked by Hook
  → MCP popup appears → Type new instructions + screenshots
  → Agent resumes with corrected context

All within the same request turn. Zero extra cost.

2. 📸 Screenshot Feedback

Stop describing bugs in words — show the AI:

| Method | Action | Description | |--------|--------|-------------| | 📷 Capture | Click button | Auto-minimizes, captures full screen | | 📋 Paste | Ctrl+V / Cmd+V | Works with system screenshot tools | | 📁 Browse | Click Browse... | Select local images |

Thumbnail preview, individual delete, auto-compress > 1600px.

3. 💬 Feedback Loop

MCP tool calls don't count as separate requests:

Agent → Popup → Your feedback → Agent continues → Popup → ...

5 iterations that would cost 5 requests → now just 1.

Architecture

┌──────────────┐    signal file    ┌──────────────────┐
│ System Tray  │ ────────────────► │ preToolUse Hook   │
│ (● click)    │                   │ (exit 2 = block)  │
└──────────────┘                   └────────┬──────────┘
                                            │ forced MCP call
                                            ▼
                                   ┌──────────────────┐
                                   │ interactive_      │
                                   │ feedback MCP      │
                                   └────────┬──────────┘
                                            │ user submits
                                            ▼
                                   ┌──────────────────┐
                                   │ afterMCPExecution │
                                   │ (clears signal)   │
                                   └──────────────────┘

Race condition protection: Two-file mechanism — cursor_interrupt (signal) + cursor_interrupt_active (marker). Clear hook only runs when marker exists.

Subagent: Can block new subagent creation. Cannot stop already-running ones.

What Gets Installed

| File | Location | Purpose | |------|----------|---------| | interrupt-check.sh/ps1 | .cursor/hooks/ | preToolUse — blocks tools on interrupt | | interrupt-check-subagent.sh/ps1 | .cursor/hooks/ | subagentStart — blocks new subagents | | clear-interrupt.sh/ps1 | .cursor/hooks/ | afterMCPExecution — clears signal | | hooks.json | .cursor/ | Hook configuration | | interrupt-hook.mdc | .cursor/rules/ | Agent behavior rule | | Tray icon | System tray | One-click trigger |

Usage

Daily Feedback

Agent auto-pops MCP at task checkpoints. Provide text + screenshots.

Emergency Interrupt

Spot the gray circle ● in menu bar / system tray
Click → "⚡ Send Interrupt"
Dot turns red 🔴
Agent's next action blocked → MCP popup
New instructions + screenshots → Submit
Dot returns gray → Agent resumes

FAQ

Q: Does interrupting waste an extra request? No. Redirects within the same request turn via preToolUse Hook.

Q: Can I interrupt during "thinking"? Not directly. Hooks fire before tool calls. But the agent will eventually make one.

Q: Already-running subagent? Cannot stop. But new creation can be blocked.

Comparison

| Feature | This project | Original | Capture | |---------|:---:|:---:|:---:| | Text feedback | ✅ | ✅ | ✅ | | Screenshot | ✅ | ❌ | ✅ | | Heartbeat | ✅ | ❌ | ✅ | | One-click interrupt | ✅ | ❌ | ❌ | | System tray | ✅ | ❌ | ❌ | | Hooks integration | ✅ | ❌ | ❌ | | One-click install | ✅ | ❌ | ❌ | | Subagent block | ✅ | ❌ | ❌ | | Cross-platform hooks | ✅ | — | — | | Predefined options | ✅ | ❌ | ✅ | | Command execution | ✅ | ✅ | ❌ |

Project Structure

cursor-snap-mcp/
├── server.py              ← MCP server
├── feedback_ui.py         ← PySide6 feedback UI
├── tray_app.py            ← System tray trigger (cross-platform)
├── hooks/                 ← Cursor Hooks
│   ├── interrupt-check.sh / .ps1
│   ├── interrupt-check-subagent.sh / .ps1
│   └── clear-interrupt.sh / .ps1
├── docs/install-guide     ← Agent-readable install instructions
├── install.sh / install.ps1
├── start-tray.sh
└── pyproject.toml

Credits

Interactive Feedback MCP by Fábio Ferreira
Capture MCP by dragonstylecc
interactive-mcp by Tommy Tong

License

MIT

MCP Servers