For the past year, Model Context Protocol (MCP) has done a lot of maturing - starting as a small open source experiment, it now became a full-fledged protocol that is broadly adopted by the industry at large. You can throw a rock and hit something that somehow integrates MCP.

That being said, one of the more peculiar limitations of MCP has always been the fact that you can’t really do a lot with it beyond text on the wire (which is still extremely useful, to be very clear) - JSON-RPC is a nice little abstraction to ferry a bunch of text-based requests and responses. Anything interactive was usually done with the help of, what I would say, hacks, like sending the URL that a user would have to go to to see a visualization of their data or results of some action (think - Playwright MCP post-test reports).

UI Comes To MCP #

Starting today, though, the landscape is changing - MCP took another leap by adding support for its first official extension, MCP Apps, a project built on the foundations of the work that the awesome folks at MCP-UI and OpenAI have been carefully nurturing.

You can catch up with my demo video to see it in action:

What’s cool about MCP Apps is that you can already tinker with it in Claude and Visual Studio Code Insiders, with more clients on the way!

Here are a few places for you to check out if you want to get building:

API Documentation Quickstart

And of course, I would be remiss if I didn’t call out the blog post announcing the release as well as the extensive collection of samples that will show you how to get started quickly.

MCP Apps are nothing other than HTML layered on top of the existing protocol abstractions, so the changes for both server and client implementers are fairly surgical. This will also hopefully make it much easier to adopt within the ecosystem.

For those reading who are a bit more security conscious, I got you covered - because apps run inside a sandboxed iframe controlled by the host, developers don’t have to worry about them escaping their container, accessing the parent page, or doing anything nefarious with cookies.

But - and that’s a big one, the real magic here is the bidirectional flow of data. An MCP App can invoke server tools and receive live updates without developers having to spin up separate infrastructure or deal with transport plumbing themselves. Neat!

A sample MCP App inside Claude AI, running on Ubuntu Linux.
A sample MCP App inside Claude AI, running on Ubuntu Linux.

Come Contribute #

The MCP Apps extension is, of course, still under active development. If you are still on the fence if you want to help shape the direction of the protocol, now is the perfect time to get involved.

For bugs or feature requests, the team is tracking everything through GitHub Issues.

For broader discussions about where the project should go or how it fits into existing workflows, there’s a dedicated space in GitHub Discussions.

And hey, if you want to take my biased opinion, this is one of those rare moments where the extension setup is still malleable enough that community feedback can genuinely shape its direction. Come help!