MCP Mac Apps Server

# Using MCP Client with Ollama ## 🎉 What is This? `mcp_client.py` - a simple Python client that combines Ollama (local LLMs) with MCP tools for managing Mac applications. **How it works:** 1. You ask a question in natural language 2. Ollama determines which MCP tool to use 3. Client calls the appropriate tool 4. Result is returned to you ## 🚀 Quick Start ### 1. Make Sure Everything is Running ```bash # Ollama should be running ollama serve # MCP server should be built npm run build ``` ### 2. Use the Client ```bash # Simple request python3 mcp_client.py "Open Calculator" # Or interactive mode python3 mcp_client.py # Then enter request ``` ## 📝 Usage Examples ### Application Management ```bash # Open application python3 mcp_client.py "Open Safari" # Close application python3 mcp_client.py "Close Calculator" # List running applications python3 mcp_client.py "What applications are currently running?" # Open file in application python3 mcp_client.py "Open file ~/Documents/test.txt in TextEdit" ``` ### Using Ollama Tools ```bash # Generate via Ollama python3 mcp_client.py "Use Ollama to explain this code" # List models python3 mcp_client.py "Show list of Ollama models" ``` ### Combined Requests You can combine requests, though the client currently processes one tool at a time. ## ⚙️ Configuration ### Changing Ollama Model In `mcp_client.py` file, change: ```python OLLAMA_MODEL = "llama3.2" # Change to your model ``` ### Changing Ollama URL ```python OLLAMA_API_URL = "http://localhost:11434" # Or other address ``` ## 🔧 How It Works Technically 1. **Getting Tool List**: Client first requests list of available tools from MCP server 2. **Creating Prompt**: System prompt is formed for Ollama with description of all available tools 3. **Querying Ollama**: Request is sent with tool descriptions and user question 4. **Parsing Response**: Response from Ollama is parsed to find JSON with tool name and arguments 5. **Calling Tool**: If JSON with tool is found, client calls the appropriate MCP tool via JSON-RPC 6. **Returning Result**: Tool execution result is returned to user ## 🐛 Troubleshooting ### Error: "Failed to Connect to Ollama" **Solution:** ```bash # Start Ollama server ollama serve ``` ### Error: "No Response from MCP Server" **Solution:** ```bash # Make sure server is built npm run build # Check that file exists ls -la dist/index.js ``` ### Model Doesn't Understand Request Try a more explicit request: - Instead of "Open calculator" → "Open application Calculator" - Instead of "Close browser" → "Close application Safari" ### Ollama Returns Text Instead of JSON This is normal for some requests. Client will return text response from Ollama if it couldn't determine a tool. ## 💡 Recommendations 1. **Use Exact Application Names**: "Calculator", "Safari", "TextEdit" (with capital letter) 2. **For Complex Tasks Use Claude Desktop**: This client is a simplified version. For more complex tasks, it's better to use a full MCP client (Claude Desktop) 3. **Check Tool List**: Client automatically gets list of available tools on each startup ## 🆚 Comparison with Claude Desktop | Feature | mcp_client.py | Claude Desktop | |---------|---------------|----------------| | Ease of Use | ⭐⭐⭐⭐⭐ | ⭐⭐⭐⭐ | | Understanding Accuracy | ⭐⭐⭐ | ⭐⭐⭐⭐⭐ | | Complex Request Support | ⭐⭐ | ⭐⭐⭐⭐⭐ | | Local Operation | ⭐⭐⭐⭐⭐ | ⭐⭐ | | Multiple Tools at Once | ❌ | ✅ | | Graphical Interface | ❌ | ✅ | ## 📚 Next Steps For more advanced usage, it's recommended to: 1. Configure Claude Desktop with MCP servers (see README.md) 2. Use multiple MCP servers simultaneously 3. Create your own more complex client with support for tool chains