> Figma has just released an MCP server, so maybe no need to work from screenshots.
could have been more clear, but composer doesn't require pasting any screenshots. the agent is able to scrape the figma directly based on the provided figma link.
the mcp wasn't a good enough integration for this and i spent hours banging my head against many approaches here. even in figma's own launch blog: the agent completely fails after a few minutes (see
video at bottom). I think this is unacceptable UX
this also doesn't solve the other massive problem which with just the figma MCP, there isn't a way for the AI to get feedback from the actual rendered version of its code (ie. imagine trying to code a UI mockup, and you can't see the rendered version of it). my work focuses on that
any feedback here would be extremely helpful