One of the key points of the comment you replied to is that you don’t need a very powerful model to do MCP stuff, such as local Qwen. The “let the agent figure it out” approach works much better with powerful models, such as Claude (which you mentioned you are using).