There's a new app store waiting to happen, and for some reason the AI labs haven't quite figured it out.
Two kinds of announcements come out of the major labs these days, especially Anthropic and OpenAI. The first is model updates. Everyone loves more intelligence for cheaper. Those are fun days.
The second kind is product-shaped, features inside the AI apps instead of model updates. OpenAI just announced a more visual interface in ChatGPT, for example. Anthropic announced motion-based artifacts. These are really cool features, and more are surely coming. But, there’s an underlying problem with an analog that goes all the way back to the original iPhone.
The iPhone launched without an app store
As many people remember, the original iPhone launched without an app store. Apple thought it would be the sole party writing whatever useful software the phone needed. Luckily, Steve Jobs was persuaded that an app store was the right approach, and it set off a boom of opportunity for software developers and users. The iPhone as a platform was too important for all the actual application development to be limited to Apple. There was no way one company could write every useful app.
We're sort of there now with the way Anthropic and OpenAI are building everything-apps with Claude and ChatGPT. They’ve added projects and artifacts and routines and code and so on. I can see the appeal of having AI apps be everything-apps, especially because AI is a different kind of platform than the iPhone was. iOS couldn't write custom code on the spot, where Claude and ChatGPT can.
But custom, on-the-fly software only gets a user so far. Expansive software takes taste, iteration, experimentation, and detail. For throwaway stuff none of this matters, but if you’re going to enhance ChatGPT or Claude with something long lasting, you have to be willing to put in the time to shape it into something great.
So there's a lot more these platforms can do than OpenAI and Anthropic are ever going to spend developer time on or the agents themselves will be able to conveniently build. There's a reason millions of apps exist in the App Store. It wasn’t only a gold rush (though it was that), or just people writing junk apps to scam people (still plenty of that). There are just a lot of different kinds of software for which the iPhone is a useful platform.
Connectors aren't apps (yet)
This is where MCP connectors come in. (MCP, the Model Context Protocol, is the open standard that lets Claude or ChatGPT talk to outside software.) Connectors aren't really applications that run inside Claude or ChatGPT. They're ways to get Claude or ChatGPT to use software that lives somewhere else. And they're amazing. They have completely changed what it's like to use AI agentically. I have dozens of connectors and use them all the time, like a lot of other people.
But unless a developer turns their connector into something you use inside Claude or ChatGPT—actual interface elements that show up as widgets in the conversation—it isn't an app in the sense that an iPhone app is.
What OpenAI recently announced with plugin extensions shows a lot of promise here. OpenAI gets teased for trying and failing to get an AI app store going, over and over. But I believe they're moving in the right direction as they keep pushing.
The reason they’ve failed up to now, other than being early and uncommitted to a strategy, is that we’ve been missing primitives an app store could be built around. We haven't had the equivalent of an SDK built on standards, at least not one the labs designed. It evolved instead: connectors, skills, and hooks, which all get bundled into plugins.
With Penny—the amazing and incredible product we’ve built that you should try right now—those pieces are enough to get widgets showing up in the chat. You can see which memory notes are being saved. You can see a list of tasks and check them off right in the widget. And we have other ideas in the works that use the same widget architecture. What's still missing is something that works alongside your conversation, not just widgets that flow by. ChatGPT plugin extensions look like a step in that direction, which is exciting.
The plugin marketplaces are a mess
Now for the part that's harder to be excited about: how miserably Anthropic and OpenAI are managing their plugin marketplaces.
Here's our experience, briefly. We're a new company with a new product, but we’ve had our product site up and the product ready to go. When we submitted a plugin to OpenAI's marketplace, OpenAI deleted our entire account. They never told us why. We assume some automated process flagged us, but nobody looked into whether we were a real company with something worth listing. We tried again and the same thing happened. We've asked, and looked, for any way to find out what happened and how to qualify. It's like talking to a black hole.
As for Anthropic, we submitted our plugin to their directory on August 26. That was more than six weeks ago. We haven't heard anything back.
Apple has been rightfully chastised at times, and criticized the rest of the time, for how it runs the App Store. Anthropic and OpenAI would face the same risks if they built a real one that turned their products into deliberate app platforms. Maybe that explains part of why they haven't.
What an app store could be
But there's room for this to be excellent. We could have MCP apps, with stores that make them discoverable, that let people improve their AI experience, beyond merely connecting to Notion or Canva or some other software that's already been built. Apps that run natively on the platforms the labs are building.
What we've built with Penny (walkthrough video below) is an opinionated and powerful way to make AI a much more useful personal assistant. It's a memory notebook that grounds your conversations with Claude or ChatGPT, so they can remember in detail what happened months ago and help you manage your to-dos and projects. We’ve built dashboard software that lives outside those apps, but it's meant to be secondary. Penny’s primary interface is Claude or ChatGPT.
Most MCP connectors work the other way around. You're meant to spend your time in a website or a software application somewhere, and it's just a nice bonus that Claude or ChatGPT can go do things there for you. It’s a bit like an iPhone only having web clipped apps. Nice, but a shadow of what’s possible.
I genuinely believe an app store is coming. But if it doesn't, I'm glad there are still ways for developers like us to get the equivalent of an MCP app installed for a user, which is how Penny works right now. It’s like the glory days of Cydia, messy and for more technical users. That at least gives us a chance to offer something we think will help a lot of people use AI more effectively. Over time, though, the right approach is a true platform.