feat: Add MCP integration to enable tools from skills.sh and other MCP servers

- Added MCP client to discover and convert MCP tools to LangChain Tools
- Extended configuration manager to support MCP server configurations
- Updated CLI to manage MCP server configurations (list, add, remove, use)
- Integrated MCP tools into the agent's tool loading system
- Added documentation for MCP integration design

docs/plans/2026-03-21-mcp-integration-design.md

@@ -0,0 +1,50 @@
+# MCP Integration Design for LangCode Agent
+
+## Section 1: MCP Client and Tool Conversion (Previously discussed and validated)
+
+## Section 2: Integration with LangCode Agent Architecture (Previously discussed and validated)
+
+## Section 3: Specific Implementation Details, Data Flow, and Error Handling
+
+### Implementation Files to Create/Modify:
+1. **New file**: `mcp_client.py` - Contains the MCPClient class and tool conversion functions
+2. **Modified file**: `main.py` - Add `get_mcp_tools()` function and integrate MCP tools into `get_custom_tools()`
+3. **Modified file**: `config_manager.py` - Add methods to manage MCP server configurations (if not already present)
+4. **Modified file**: `config_cli.py` - Add CLI commands for MCP server management (list, add, remove, use, etc.)
+
+### Data Flow:
+1. **Startup**: When the agent initializes, `get_custom_tools()` is called
+2. **MCP Tool Discovery**: `get_mcp_tools()` reads all configured MCP servers from config_manager
+3. **Server Connection**: For each MCP server, an MCPClient instance connects to the server URL
+4. **Tool Enumeration**: The client calls the MCP server's `tools/list` endpoint to get available tools
+5. **Tool Conversion**: Each tool specification is converted to a LangChain Tool with:
+   - Same name and description as the MCP tool
+   - A function that serializes arguments and calls the MCP server's `tools/call` endpoint
+   - Proper error handling for network issues, server errors, or invalid responses
+6. **Tool Aggregation**: All MCP-derived tools are combined with existing Proxmox and provider tools
+7. **Agent Initialization**: The DeepAgent is created with the complete toolset
+
+### Error Handling Strategy:
+- **Connection Failures**: If an MCP server is unreachable during startup, log a warning and continue (don't fail agent initialization)
+- **Tool Call Failures**: If a tool call to an MCP server fails, return a descriptive error message as the tool result
+- **Schema Validation**: Validate tool arguments against the MCP tool's input schema before making the call
+- **Timeouts**: Implement reasonable timeouts for MCP server connections and calls
+- **Retry Logic**: Consider implementing retry logic for transient network failures
+
+### Testing Approach:
+1. **Unit Tests**: 
+   - Test MCPClient connection and tool listing with mock servers
+   - Test tool conversion function with various tool specifications
+   - Test error handling scenarios (timeouts, invalid responses, etc.)
+
+2. **Integration Tests**:
+   - Test with a real MCP server (could be a local test server or skills.sh if accessible)
+   - Verify that tools appear in the agent's tool list and can be invoked successfully
+   - Test end-to-end flows: user asks agent to use an MCP-derived tool, agent executes it correctly
+
+3. **Manual Testing**:
+   - Add an MCP server configuration pointing to skills.sh (if it offers MCP endpoints)
+   - Verify that skills from skills.sh appear as available tools in the agent
+   - Test invoking a few different skills to ensure they work correctly
+
+Are these implementation details, data flow, error handling, and testing approaches correct and complete? Do you see any gaps or areas that need adjustment?