The Last User Interface

A few months ago I predicted that by 2027 most GUIs would be replaced with conversational AIs. I would like to make two amendments:

  1. I predicted that each product or service would have its own LLM. However, a more likely way forward is that services will implement the Model Context Protocol, and agentic LLMs will use this to control them.
  2. The move from GUIs to conversational AI will happen sooner than 2027, at least for utility applications. “Specialist” applications, like image manipulation software, video editing, word processors, etc will still be operated manually. However, email clients, todo apps, calendars etc will be used primarily through LLMs.

To give an example use case, imagine that you have a handwritten to-do list. You could then take a photo of it, and ask your LLM to add the to-do items to an MCP-enabled To-do application. The To-do application doesn’t even need AI capabilities, it just needs to expose an API that’s easy for an LLM to interact with. I think a lot of applications will follow this pattern—you will have a single “smart”, multimodal LLM that can interact with a multitude of “dumb”, specialized tools.

Another advantage of an MCP interface is that they can be made a lot more sprawling than would be possible with a GUI, since a GUI is by necessity constrained by what can be displayed on a screen. In practice, it is even more constrained since the interaction has to make sense to the user—you can’t just fill the UI with buttons and expect the user to understand how it works. MCP doesn’t have that limitation—this means that use cases that would be too niche to have support in a GUI (“give me the top five to-do items that are due today, sorted by due date”) would be possible with MCP.

A secondary prediction is that there will soon be a cottage industry for MCP implementations; MCP developers might be the 2020s equivalent of late 90s web developers.

Speaking of web, will people even build websites any more? Or will they just be another MCP implementation? And how will you find these websites? The MCP web? Maybe general-purpose LLMs will replace the web browsers as the primary way we interact with businesses and their services?

Imagine that you want to book a flight ticket; you tell your LLM where you would like to go, and when. You can also add any other constraints you want (price, airline, maximum flight time)—your “browser” LLM will contact an MCP search engine to find MCP-enabled travel agencies. It will then ask each of them for a list of flights that match your criteria, and present you with the best options. Given your approval, it will book one of them for you, and add an entry to your calendar. When the day of your flight comes, your LLM will be able to fetch your boarding pass—no need to fuss about with emails, PDFs or an Apple Wallet. If the LLM has a voice mode there’s no need to even look at a screen during the entire process; the entire user interface is just a conversation.