DesktopVisionMCP. Show, don't tell.

You describe your screen to your assistant. You screenshot it, crop it, paste it, and still leave out the thing that mattered. DesktopVisionMCP lets it look instead. A macOS menu bar app and a plain MCP server, so it works with Claude, ChatGPT, Codex, Gemini, or whatever you move to next. It captures only when asked. It never records, never clicks, never types, and its state is always in the menu bar. Captures go to the assistant that asked, and nowhere else.