Add MCP (Model Control Protocol) Client Support

[wendy-aw]

Background

Model Context Protocol provides a standardized way for LLMs to use external tools. This PR adds support for connecting to MCP servers and letting models use tools working with both Anthropic, OpenAI and Gemini models.

Changes

• Add a complete MCP client implementation for connecting to both SSE and stdio MCP servers
• Add comprehensive testing suite with live API and server connection tests
• Integrate with current Anthropic, OpenAI and Gemini API handling

How to Use MCPClient

from defog.llm.utils_mcp import initialize_mcp_client


# Initialize with config file
client = await initialize_mcp_client("path/to/config.json",
"claude-3-haiku")

# Chat with tool usage
response, tool_outputs = await client.mcp_chat("Use tools to help me
   multiply 12 and 34")

# Clean up
await client.cleanup()

[rishsriv]

Thank you! This is awesome, and I'm super excited to use it in more of our apps!

Just 1 request – looks good to merge after that is changed!

[rishsriv]

Instead of the first input param being a config_path, let's rename it config – where the type can either be dict or a string (filepath)

With this, calling this function from other apps becomes significantly more flexible. We can either pass a dict (which can be programmatically generated) to the function, or we can pass a string.

[wendy-aw]

Ok done! Lemme know if this is ok!

[rishsriv]

Thank you!