The LLM passes the HTML back in their response, translated, so it can be rendered to the user as-is. The tool description is self-explanatory: https://github.co

The LLM passes the HTML back in their response, translated, so it can be rendered to the user as-is. The tool description is self-explanatory: https://github.com/TuanKiri/weather-mcp-server/blob/master/internal/server/tools/weather.go#L13-L19 https://github.com/TuanKiri/weather-mcp-server/blob/master/i... You could have it respond with JSON, but then the rendering of the weather widget would have to be defined by the consumer instead of the tool server -- now the 'chat' part of your system needs to be aware of all possible content types, how to detect and render them; it's also doable but a different set of trade-offs.