I get where you’re coming from, but there are some security practices in place. The host client renders views inside a strictly sandboxed `<iframe>`. Any action the UI wants to take must pass an auditable message back to the host application, which triggers an explicit user-permission prompt.
I stumbled onto this capability while using the Spotify MCP App inside Claude. a fully functional, interactive playlist an dmusic player widget spin up inline, instead of a standard markdown list of text links. It turns out this is built on the official MCP Apps spec extension that Anthropic and OpenAI standardized a few months back. It lets your server declare a ui:// URI, which the host grabs and mounts inside a sandboxed iframee. Core mcp blew up overnight, but this extension still feels pretty underground. It makes me wonder if we're heading toward a weird shift where AI agents become the primary browsers of the web, and front-end dev becomes about building micro-widgets for LLMs instead of humans.