I want to build an agent/assistant that has no default internal tools. Everything should be an MCP tool, allowing maximum flexibility in configuring an environment.
In agent mode this is already possible, but in assistant mode I'm lacking a key feature: preview.

In my previous implementation, while I supported MCP, I also provided internal tools for things like file editing. There were two key reasons for this that I'm hoping to solve with preview mode:

• I want to inform the user contextually of what is about to happen before he clicks accept/reject. Just showing the raw input of a tool call is not enough.
• I want to do a preflight to make sure that the tool call makes sense before prompting the user to reduce interaction. This means validation beyond the simplistic "does it match the parameter spec".

I propose:

• we use capabilities on a tool call to indicate that preview mode is available (e.g. preview: true)
• we also define an optional preview ui endpoint, (e.g. previewResource: ui://preview/{previewId}). If no endpoint is given, the input/output of the tool call is deemed sufficient to inform the user.
• we use a new _meta field to indicate that the tool call should be run in preview mode (preview: true)
• the tool call result should return a previewId in the _meta section (perhaps also preview: true in case no resources are used)

When preview is toggled at runtime, the MCP server does a dry-run of the call, provides (as much as possible) the same output as it would when actually running and optionally uses a resource to visually inform the user.

This means:

• the input/output spec of the tool call remains exactly the same (so no variable output due to preview)
• preview mode can be used both as a preflight check and a way to generate a contextually relevant view
• by allowing for previewId, we can support both "static" previews and "dynamic" previews. I would suggest that the tool input/output is always fed into the iframe BUT the preview can also use internal knowledge that is not easily recreated from the input/output or the MCP bridge by generating a dynamic resource.
• the choice of previewing or not is entirely in the client application. Neither the LLM nor the MCP server need to manage this.
• in agent mode (or if the user clicks "allow all" in assistant mode) you can simply call the tool without the preview toggle