After playing with mcp for a few days I just built my own MCP server. It uses GPT-4o + embeddings to understand the screenshot and let them search naturally. It's published on pypi and works with claude desktop.
Mcp is a cool thing. Just connected an agent with two tools using mcp and gave it a task.
Open vaibhav kothari GitHub
Find his work on llm hallucinations
Open the readme
Create a summary and save it in sandbox folder.
And it took me like 10 minutes