MCPDesk

A Windows-native desktop client for Model Context Protocol (MCP). Connect to any MCP server, use any LLM provider, and orchestrate AI tools — all from one app.

Why MCPDesk?

Every existing MCP client is either locked to one LLM (Claude Desktop) or is a code editor (VS Code, Cursor, Zed). MCPDesk fills the gap:

Feature	MCPDesk	Claude Desktop	VS Code + Continue	Cursor
Multi-LLM provider	✅	❌ Claude only	✅	✅
GitHub Copilot support	✅	❌	✅	❌
Non-developer friendly	✅	✅	❌	❌
Windows native (WinUI 3)	✅	❌ Electron	❌	❌ Electron
Multi-MCP server	✅	✅	✅	Limited
Mica backdrop	✅	❌	❌	❌

Features

• Multi-Provider LLM Support — OpenAI, Anthropic Claude, GitHub Copilot, and Ollama (local). Switch providers and models on the fly.
• GitHub Copilot Integration — Sign in with your existing Copilot subscription via OAuth device flow. No extra API keys needed.
• Dynamic Model Selector — Browse all models available in your subscription and switch instantly.
• MCP Server Management — Configure, connect, and manage multiple MCP servers. Auto-connects enabled servers on startup.
• Tool Call Orchestration — The LLM sees tools from ALL connected MCP servers. Multi-iteration tool calling with real-time status updates.
• Conversation Persistence — SQLite-backed chat history with conversation management (create, switch, delete).
• Windows 11 Native — WinUI 3 with Mica backdrop, NavigationView, and proper Windows design language.
• Config Auto-Detection — Paste standard MCP config formats and MCPDesk normalizes them automatically.

Quick Start

Prerequisites

• Windows 10 (1809+) or Windows 11
• .NET 8 SDK
• Windows App SDK

Build & Run

git clone https://github.com/clatonhendricks/MCPDeskClient.git
cd MCPDeskClient
dotnet restore src/MCPClient/MCPClient.csproj
dotnet build src/MCPClient/MCPClient.csproj
dotnet run --project src/MCPClient/MCPClient.csproj

Configuration

MCPDesk stores its configuration at %APPDATA%\MCPDesk\config.json:

{
  "mcpServers": {
    "kusto-mcp": {
      "command": "uvx",
      "args": ["azure-kusto-mcp"],
      "env": {
        "KUSTO_SERVICE_URI": "https://your-cluster.kusto.windows.net/"
      },
      "enabled": true
    },
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "C:\\Users\\you\\Documents"],
      "env": {},
      "enabled": true
    }
  },
  "llmProviders": {
    "github-copilot": {
      "type": "GitHubCopilot",
      "apiKey": "",
      "model": "gpt-4o",
      "enabled": true
    }
  },
  "defaultProviderId": "github-copilot"
}

GitHub Copilot Setup

1. Open MCPDesk → Settings
2. Click Sign in with GitHub
3. Enter the code shown at github.com/login/device
4. Done — your Copilot subscription is now the LLM backend

Adding MCP Servers

Via config.json — Edit %APPDATA%\MCPDesk\config.json and add entries under mcpServers.

Via UI — Navigate to MCP Servers → Add Server → fill in command, args, and environment variables.

Architecture

MCPClient/
├── src/
│   ├── MCPClient/                    # WinUI 3 app
│   │   ├── Views/                    # ChatPage, SettingsPage, ServerConfigPage
│   │   ├── ViewModels/               # MVVM with CommunityToolkit.Mvvm
│   │   ├── Converters/               # XAML value converters
│   │   └── App.xaml                  # DI container, global error handling
│   └── MCPClient.Core/               # Business logic (no UI dependency)
│       ├── LlmProviders/             # OpenAI, Anthropic, Ollama, GitHubCopilot
│       ├── Services/                 # MCP client, config, LLM, conversation
│       └── Models/                   # AppConfig, Conversation, Message
└── tests/
    └── ToolCallTest/                 # Integration test for tool call loop

Key Technical Details

• Copilot API: OAuth tokens are exchanged for Copilot JWT tokens via api.github.com/copilot_internal/v2/token. Chat uses direct HTTP to {endpoint}/chat/completions (not OpenAI SDK, which prepends /v1/).
• Tool Names: MCP tool names use __ separator (server__tool) to stay within OpenAI's [a-zA-Z0-9_-] function name constraint.
• Provider Persistence: Provider instances are reused across page navigations to preserve auth state.

Supported LLM Providers

Provider	Auth Method	Models
GitHub Copilot	OAuth device flow or PAT	All subscription models (GPT-4o, Claude, Gemini, etc.)
OpenAI	API key	GPT-4o, GPT-4o mini, o3-mini
Anthropic	API key	Claude Sonnet 4, Claude Opus 4, Claude Haiku 3
Ollama	Local endpoint	Any locally installed model

Dependencies

• Microsoft.WindowsAppSDK — WinUI 3
• CommunityToolkit.Mvvm — MVVM framework
• ModelContextProtocol — MCP C# SDK
• Microsoft.EntityFrameworkCore.Sqlite — Conversation persistence
• OpenAI — OpenAI/GitHub Models SDK
• Anthropic.SDK — Anthropic SDK

Roadmap

• MCP Server Gallery — Browse and one-click install popular MCP servers
• Streaming responses — Show tokens as they arrive
• Workflow/Agent mode — Save repeatable tool chains
• Export to Markdown/PDF
• MSIX packaging for enterprise deployment
• Conversation search
• System tray / hotkey quick launch
• Markdown rendering in chat

License

GPL-3.0