AI代理对话系统

Consult User MCP

Native dialog system for MCP (Model Context Protocol) servers. macOS and Windows.

Keep working while your AI agent runs autonomously. When it needs your input, a native dialog appears - answer the question, and the agent continues. No need to watch the terminal. Snooze if you're busy, or send feedback to redirect the agent.

See screenshots and full documentation →

Quick Install

macOS:

curl -sSL https://raw.githubusercontent.com/doublej/consult-user-mcp/main/install.sh | bash

Windows: Download the installer from Releases and run it. First-run wizard automatically configures Claude Code.

Contents

• Install
  • macOS
  • Windows
  • MCP Server Only
• Uninstall
  • macOS uninstaller
  • Windows uninstaller
• Build from Source
• Structure
• MCP Tools
• Architecture

Install

macOS

Quick install:

curl -sSL https://raw.githubusercontent.com/doublej/consult-user-mcp/main/install.sh | bash

Manual install:

1. Download Consult.User.MCP.dmg from Releases (look for macos/vX.Y.Z)
2. Open it and drag the app to /Applications
3. Launch the app - a menu bar icon appears
4. Click the menu bar icon and use Install for Claude Code to automatically configure the MCP server

The app is signed with a Developer ID and notarized by Apple, so there is no quarantine flag to clear. The .zip asset beside the dmg is what install.sh and the in-app updater consume.

Windows

1. Download the installer from Releases (look for windows/vX.Y.Z)
2. Run the installer - it configures everything automatically
3. The system tray app launches at startup

The first-run wizard automatically configures Claude Code. Settings, snooze state, and history are stored in %APPDATA%\ConsultUserMCP\.

MCP Server Only

If you don't want the tray app, you can run the MCP server standalone:

git clone https://github.com/doublej/consult-user-mcp.git
cd consult-user-mcp/mcp-server
bun install && bun run build

Add to your MCP config:

{
  "mcpServers": {
    "consult-user-mcp": {
      "command": "node",
      "args": ["/path/to/consult-user-mcp/mcp-server/dist/index.js"]
    }
  }
}

Uninstall

macOS uninstaller

curl -sSL https://raw.githubusercontent.com/doublej/consult-user-mcp/main/uninstall.sh | bash

Optional flags:

• --keep-data keeps ~/Library/Application Support/ConsultUserMCP and ~/.config/consult-user-mcp/projects.json
• --dry-run prints planned changes without modifying files

Windows uninstaller

You can uninstall from Installed apps in Windows settings.

For full cleanup (MCP entries + base prompt block + app files), run:

powershell -ExecutionPolicy Bypass -File .\uninstall.ps1

Optional flags:

• -KeepData keeps %APPDATA%\ConsultUserMCP and %USERPROFILE%\.config\consult-user-mcp\projects.json
• -DryRun prints planned changes without modifying files

Build from Source

macOS:

bun install
bun run build:bundle   # Builds and installs to /Applications

Windows: See CLAUDE.md for Windows build instructions (requires .NET SDK 8.0 and Velopack).

Structure

consult-user-mcp/
├── mcp-server/          # MCP server (TypeScript, cross-platform)
├── dialog-cli/          # macOS: Swift CLI for dialogs
├── sketch-cli/          # macOS: Swift CLI for layout editor
├── macos-app/           # macOS: SwiftUI menu bar app
├── dialog-cli-windows/  # Windows: WPF CLI for dialogs
├── windows-app/         # Windows: WPF system tray app
├── docs/                # Documentation site (SvelteKit)
└── scripts/             # Build and release scripts

MCP Tools

Timeout and Snooze

All interactive dialogs (ask, tweak) have a 10-minute timeout.

Snooze: Users can snooze a dialog for 1-60 minutes. When snoozed:

• Returns { snoozed: true, remainingSeconds: S }
• All subsequent calls return the same until snooze expires
• Agent should sleep for remainingSeconds, then retry

Feedback: Users can send text feedback instead of answering. Returns { feedbackText: "..." } for the agent to adjust its approach.

Architecture

flowchart TB
    subgraph Agent["AI Agent (Claude Code, etc.)"]
        A[Tool Call]
    end

    subgraph MCP["MCP Server (Node.js)"]
        B[index.ts]
        C{Platform?}
        D[SwiftDialogProvider]
        E[WindowsDialogProvider]
    end

    subgraph macOS["macOS"]
        F[dialog-cli Swift]
        G[SwiftUI Window]
    end

    subgraph Windows["Windows"]
        H[dialog-cli.exe WPF]
        I[WPF Window]
    end

    subgraph User["User Response"]
        J[Answer]
        K[Snooze]
        L[Feedback]
    end

    A -->|"stdio (JSON-RPC)"| B
    B --> C
    C -->|"darwin"| D
    C -->|"win32"| E
    D -->|"execFile"| F
    E -->|"execFile"| H
    F --> G
    H --> I
    G --> J & K & L
    I --> J & K & L
    J & K & L -->|"JSON stdout"| B
    B -->|"stdio"| A

    style Agent fill:#e1f5fe
    style MCP fill:#fff3e0
    style macOS fill:#f3e5f5
    style Windows fill:#e3f2fd
    style User fill:#e8f5e9

Flow

1. Agent calls tool - Claude Code (or any MCP client) invokes ask, notify, or tweak
2. MCP Server receives - TypeScript server validates input with Zod schemas
3. Platform dispatch - Server spawns the native CLI for the current platform
4. Dialog shown - Native window appears (SwiftUI on macOS, WPF on Windows)
5. User responds - Answer, snooze, or feedback
6. Result returned - JSON flows back through CLI → MCP Server → Agent