Loading video...
Video Failed to Load
OpenAI, Anthropic and Microsoft all back the same standard for running apps inside AI chats. Booking .com, Figma, and Canva are already using it. Still, most developers have never heard of it. It's called MCP Apps. So, why does it matter? The next time someone buys from your website,... show more
15,470 views • 2 days ago •via X (Twitter)
23 Comments
When an MCP App renders inside the chat, does it inherit the host's tool grants or get its own? That one answer decides most of the security story. Booking and Figma are easy cases. The hard one is an app calling a tool the user never connected.
how does auth work when the ui renders inside the chat. does the user log in through the iframe or is it passed from the host
Good question. The host coordinates OAuth, usually opening the provider’s login page outside the iframe. It then sends the access token only to the MCP server. The embedded UI calls authenticated tools through the host, so the token itself is not exposed to the UI.
Does the same MCP App render the same in ChatGPT and Claude, or do you end up maintaining per-host quirks like we did with browsers?
Great question. The same React view can run in both, but pixel-perfect and capability-perfect parity is not guaranteed. Hosts differ in layout modes, styling, permissions, and supported APIs. Skybridge hides the protocol differences, while apps still need capability checks and graceful fallbacks.
the clicked state is the part most MCP demos miss
The key is not the click itself, but the state transition it represents. If “add to cart” succeeds, the app must sync the confirmed cart state back to the model. Otherwise the UI and conversation become two conflicting sources of truth.
the part i'd want to see tested is error states, what happens when the cart sync breaks mid chat
That’s the real production test. The server should remain the source of truth, with failed tool calls surfaced in the UI and a retry or refresh path. Model context should only reflect confirmed cart state, never an optimistic click.
@akshay_pachaar interesting how MCP Apps might reshape things, but the real challenge for devs is integrating these new standards without disrupting existing workflows.
i checked about 50 companies in perplexity answers over the last two weeks. more than half never got named... showing up as an app comes after that
How does checkout work inside the chat? I'd need to see that before trusting it with my card.
Apps inside chats only help if the product data underneath is clean first. Accurate price, stock and variants in structured data decide whether an agent can surface an item at all, before any checkout flow matters.
MCP Apps make interactive product flows easier to bring into chat, shifting the hard part from presenting the UI to making its state and actions trustworthy.
already living inside the chat is the part that matters for shops. Curious if most MCP apps youve seen finish the job in-thread, or still dump people into a tab halfway.
MCP Apps is the quiet part: one UI bridge, three labs, Booking/Figma/Canva already in-chat.#MCP #AIAgents @akshay_pachaar @OpenAI @AnthropicAI
Years of A/B testing the checkout, and now the render happens in someone else's chat window. Getting a model and a live UI to agree about what's in the cart is the part that hurts.
MCP Apps are going to introduce an entirely new class of proxy caching bugs.
Consider exploring how MCP apps display interactive user interfaces, such as cards or add-to-cart flows, directly inside AI chats.
a shared standard for apps inside chats is what makes it worth building for. one integration, every assistant
This changes the design brief completely. We used to design pages, now we're designing a card the model decides when to show. Who owns the brand experience then?
MCP Apps is the right shape for this. the part nobody talks about: once the app runs inside the chat, the payment rail has to live there too. UI in the conversation, settlement underneath. that's the x402 gap rn.
The pace of innovation is incredible.
