MCP Client-Server Demo — XCodeFreeze

XCodeFreeze is a macOS SwiftUI chat app that works as a Model Context Protocol (MCP) client and an AI front-end at the same time. It starts an MCP server (by default the xcf Xcode MCP server) over stdio, finds the server's tools, and lets a local LLM call those tools during a chat. The LLM suggests a tool call, the client runs it, and the result goes back to the model.

Built on the official MCP Swift SDK (0.8.2, up to next major).

Features

• MCP over stdio: starts the server process and connects with StdioTransport on FileDescriptors. It can connect and disconnect cleanly and send raw messages for debugging.
• Tool discovery and registry: lists the server's tools and their JSON-schema parameters, and keeps them in a type-safe registry.
• AI chat with tool calling: talks to any OpenAI-compatible endpoint (/v1/models, /v1/chat/completions), such as Ollama on port 11434. MCP tools are passed to the model as function tools. Tool calls are run and the results are fed back to the model, several rounds if needed.
• Thinking: shows the model's <thinking> blocks separately from the answer.
• Model picker: fills the model list from /v1/models. The LLM server address can be changed in the header.
• Config picker: choose an existing MCP JSON config or create a new one. The path is saved in UserDefaults.
• Built-in commands: help, list, ping, plus Debug and Diagnostics buttons.

Configuration

Click New in the header to create a config file (mcp.json):

{
  "mcpServers": {
    "XCF_MCP_SERVER": {
      "type": "stdio",
      "command": "/usr/local/bin/xcf",
      "args": [],
      "env": {}
    }
  }
}

Set command to the path of your MCP server executable. The app uses the first server in the file (fallback name XCF_MCP_SERVER).

The LLM server address defaults to 192.168.1.135:11434. Change it in the app header to your own Ollama / OpenAI-compatible host.

Project layout

XCodeFreeze/
├── XCodeFreeze.xcodeproj
└── XCodeFreeze/
    ├── Config/     MCPConstants, ServerConfig
    ├── Models/     MCPViewModel, ChatMessage, AIModels (OpenAI-style request/response types)
    ├── Services/   TransportService, ClientServerService, MCPCommunicationService,
    │               MCPToolDiscoveryService / ToolDiscoveryService, ToolRegistry,
    │               AIService, MCPAIIntegrationService, CommandService,
    │               ConfigFileService, LoggingService, JSONFormatter, MessageFormatService
    └── Views/      ContentView, MessageView, Components/ (JSONHeaderView, ChatView,
                    InputBarView, ToolsListView, AIModelSelectorView, ConfigSelectorView)

XCodeFreeze/README.md has notes on how the SDK is integrated (FileDescriptor transport, disconnect handling, raw-message debugging).

Build and run

1. Open XCodeFreeze/XCodeFreeze.xcodeproj in Xcode and run the XCodeFreeze scheme.
2. Select or New an MCP config, then Connect.
3. Set your LLM server address, pick a model, and start chatting.

History

• May 2025: started as a basic MCP client/server demo (transport, JSON handling, enums, type-safe tool registry) and was renamed XCodeFreeze.
• Aug–Sep 2025: added AI chat. The model suggests tools and the client runs them. Added the single-line JSON header UI and fixed duplicate logging and over-eager tool auto-execution.