Add MCP server tool loading to Agent SDK

[flyworker]

Context

mega-nova-api is implementing an MCP server (MeganovaAI/mega-nova-api#437) and nova_mesh already has MCP client support. The meganova-python Agent SDK should be able to load tools from MCP servers.

Current Agent SDK Tool System

@tool("get_weather", "Get weather")
def get_weather(city: str) -> str: ...

agent = Agent(client=client, tools=[get_weather])

Tools are registered via @tool decorator → ToolRegistry → OpenAI JSON Schema.

Requirements

• Add optional mcp dependency: pip install meganova[mcp]
• Create meganova/agents/mcp.py with MCPToolLoader
• Support loading tools from MCP servers:

  from meganova.agents.mcp import load_mcp_tools
  
  mcp_tools = load_mcp_tools("npx @meganova/mcp-server")
  agent = Agent(client=client, tools=[get_weather, *mcp_tools])

• Support STDIO and HTTP/SSE transports
• Convert MCP tool schemas to ToolDefinition format
• Handle MCP tool execution through MCP client protocol
• Add example: examples/agent_with_mcp.py

Dependencies

• mcp Python SDK (optional extra)

Reference

• nova_mesh implementation: nova_mesh/mcp/client.py, nova_mesh/mcp/skill_pack.py

[flyworker]

📋 Reference Report: AI API Compatibility Best Practices Research (29 items, 7 categories — protocols, gateways, SDKs, cloud platforms, local inference)

Tracking Epic: MeganovaAI/mega-nova-api#447