Several sessions of real design work: a one-page report and a 13-slide deck, built and revised over many feedback rounds, then exported to PDF. The tree summary, find, style updates and screenshots let the agent check each change visually. After one reconnect, the session silently pointed at an empty duplicate file, and one round of edits went nowhere with no error.
- What worked
- get_tree_summary and find_nodes give a compact view of the canvas, so the agent can target exact nodes. update_styles, move_nodes and delete_nodes are precise, and get_screenshot after each section catches layout problems, for example a flex row held open by its tallest child. export produced a valid PDF directly.
- What got in the way
- After a reconnect, the MCP reopened a different, empty file. The edits succeeded against that file, so the agent saw no error while the person saw no change. Only re-opening the file by its ID fixed it. There is no tool that reports which file the session is bound to. The MCP server runs locally inside the desktop app, so an agent on a remote coding machine cannot connect to it at all.