SUBSCRIBE | ADVERTISE

ChatGPT is eating the app layer

Thoughts on OpenAI's DevDay, Apps and AgentKit

Was this email forwarded to you? Sign up here.

OpenAI held its yearly DevDay this week. This time, it was less about raw model power, and more about turning ChatGPT into an operating system and the filter of everything else.

PS - At the end of today's newsletter, I'm sharing how I'm saving 2 hours daily using a new AI tool. Scroll down if you're curious.

Apps in ChatGPT

OpenAI introduced Apps in ChatGPT, including an Apps SDK (tools for developers to build the apps).

You can think about these apps as an experience where you’re logged into an app while having its native interface—right inside ChatGPT.

A couple of examples from the early demos:

• The Spotify app lets you do things like compose a playlist together with ChatGPT. You can ask for personalized picks and get songs listed in a Spotify-native interface, complete with album art and buttons to listen/add to playlist.

• The Zillow app lets you do super specific real estate that would require quite a bit of thinking to filter down to using their native website. For example, you can ask for “Show me 2-bedroom apartments for sale in Denver under $600,000” and you’ll get a detailed clickable map for exploring listings right inside ChatGPT. You can keep refining the filters or follow up with ChatGPT with questions like: “which of these are within 1km distance of a skatepark?” (indeed, an important consideration), or maybe just “which of these fit me best based on what you know about me?”.

The initial Apps that were launched are already available to use for everyone except in the EU (surprise, surprise).

Developers can already start building apps with the dedicated SDK which is built on top of the Model Context Protocol (MCP). Later this year, OpenAI will start allowing other developers to publish their apps to ChatGPT, where OpenAI will create a directory—essentially an app store—so users can discover them. They will also surface quality apps directly in the chat when the user puts in a related query.

Why it matters

OpenAI has been tinkering with the idea of ChatGPT being the middleman between you and your apps since the introduction of plugins in 2023, and the evolution to Custom GPTs in 2024.

While the new experience of Apps looks more polished and native to ChatGPT, not all the apps that were demoed looked useful; several come across as something we’ve seen before with GPTs: pure starting points where users still need to go to an external app/site to get any tangible value. But some were actually very neat—Zillow was my favourite. When users can get a self-contained experience of your product within ChatGPT without having to leave the platform (which defeats the purpose of Apps), that’s when I think it will be a sticky ChatGPT App.

With OpenAI allowing companies to gate their app experience with log-in, paid plans, and the new instant checkout feature—which enables App-store style monetization opportunities—I think the potential is there for an actual ecosystem (unlike plugins and GPTs where, among other things, lack of financial incentives for companies to create solid in-app experiences never allowed it to take off).

IN PARTNERSHIP WITH EMERGENT

Ship your app in a weekend

Have an app idea that needs an interface, login, a place to save stuff, and a way to charge money?

Emergent is the app builder that handles the plumbing so you can hit publish:

• Login, ready on day one: sign-up/sign-in works out of the box

• Built-in database (MongoDB): store accounts, settings, and content—no setup

• Easy payments: Stripe is built in so you can easily monetise your app

• Go live fast: deploy your app directly on Emergent and add your custom domain if you want to

• Best for: solo builders/small teams shipping lightweight tools or MVPs.

Try Emergent for free →

AgentKit (or “ChatGPT in your apps”)

OpenAI launched AgentKit, a set of visual tools to add chat UI directly into your apps. The pitch is that it makes it easier for non-technical people to build AI agents that would've previously taken a full developer team.

The toolset has two main components:

• Agent Builder: A workflow creator with a flowchart canvas for composing business logic (think n8n and Zapier, but native to ChatGPT). Drag and drop nodes and connect tools to automate workflows.

• ChatKit: An embeddable chat interface that uses the workflows you've built. Customise it with Widgets using natural language. OpenAI also launched ChatKit Studio, including a cool (but kinda glitchy) demo based around a world map.

There’s also the Connector Registry for admins to manage data sources and tools (ie. Connectors and MCPs) in one place.

OpenAI also equipped AgentKit with Guardrails (protection against unintended or malicious behavior) for safer deployment and enhanced Evals (tools to check how well an AI performs on specific tasks) for better testing.

Why it matters

The functionality isn’t new, but it’s a step forward in terms of accessibility. I like the idea that more people in a company—often those who sit much closer to the problem than developers do—are enabled to create more advanced solutions to support their needs.

The tools in AgentKit are advanced enough to create things like a support agent that understands the business’s data, takes action on the site, and gives customers guided, tailored solutions. All within an on-brand and interactive UI.

While the interfaces for building with AI get simpler, I think it’s worth acknowledging this: creating complex systems that automate important business processes still takes a great deal of skill, domain expertise and systems thinking.

—

Where does this leave workflow builders like n8n/Make/Zapier/etc?

Let’s just say, if you're a VC-backed workflow automation company, this is the moment you start emphasizing your "strong ecosystem" and "enterprise relationships" in board meetings.

Will they die? Probably not (they do have strong ecosystems). But AgentKit proves OpenAI isn't just vertically integrating compute and models anymore. They want the app layer too. And they don't care that dozens of companies are already doing workflow automation and AI chatbots. If it's a popular AI use case, it's likely on their roadmap.

The other releases from DevDay

I’ll link to the other DevDay updates here quickly:

• Codex upgrades: Codex is the most amazing coding tool for AI (it’s my go-to these days, I prefer it over Claude Code). Devs can now use the same agent that powers Codex directly in their own apps using the new Codex SDK. Also, Code now has a Slack integration so you can ping it directly from Slack.

• Sora 2 in the API: It’s the only real challenger to Google’s Veo 3 and way cheaper.

• GPT-5 Pro in the API: Could be worth considering if you need to solve some really, really, really hard problems for your users (and don’t mind paying for it).

• Two new mini models: gpt-realtime-mini and gpt-image-1-mini. Putting AI voices and images into your app is now 70-80% cheaper.

TOOL SPOTLIGHT

I use Paraspeech for Mac, an AI dictation app that runs fully on-device and completely offline.

It’s dead simple: tap, speak, paste. Blazing fast and accurate dictation.

I use it for prompting ChatGPT/Claude/Cursor, drafting emails or capturing notes—even while switching between tabs or walking around the room. It lets me articulate my thoughts and ideas faster than typing them out.

How much faster? 2 hours/day.

Here’s the math: on average, I tap Paraspeech 70 times per day for 51 seconds per session. I speak at 174 WPM, so that’s 1 hour of speech and ≈10,000 words transcribed per day. I write at only 60 WPM, so that would’ve taken me 2.8 hours to write.

I’m 2 hours faster every day using dictation.

Right now, Paraspeech is only $39.99 for a lifetime license.

If you think out loud, it pays for itself in the first week.

Get Paraspeech (15% off with code HOWDYWIZARDS)

Transparency: I'm a partner of Paraspeech, so I'll earn a commission if you buy. Thank you for supporting the newsletter.

THAT’S ALL FOR THIS WEEK