OpenAI Wants the Customer Before Your App Gets the Click
What happens to your app when the customer never needs to know its name?
That’s the uncomfortable question behind OpenAI’s September 29 DevDay announcements. Not whether ChatGPT can display another widget. Whether it can become the place where people describe a problem, software gets selected, and an agent does the work.
I’m excited about that interface. I’m considerably less excited about who controls the recommendations.
The storefront is becoming a conversation
According to [OpenAI’s DevDay recap](https://openai.com/index/devday-2026-recap/), the platform is adding dedicated plugin sidebar locations, interactive panels, supported file viewers, and improved discovery. Supported plugins can also appear in shared ChatGPT Sites, with teammates using their own connected data and permissions.
These aren’t isolated UI upgrades. Together, they make software usable without sending the customer elsewhere.
The groundwork started on October 6, 2025, when OpenAI introduced the Apps SDK, built on Model Context Protocol (MCP), with seven partners: Booking.com, Canva, Coursera, Expedia, Figma, Spotify, and Zillow. Conversational recommendations were already part of that pitch.
What’s new now is the surrounding machinery:
- Identity: Sign in with ChatGPT.
- Portable usage: eligible Plus and Pro subscribers can use their ChatGPT allowance in participating external tools.
- Automation: support for the proposed MCP Events specification.
- Procurement: an enterprise marketplace launching with 32 partners.
That’s a distribution stack taking shape. Not a finished app-store replacement.
<> OpenAI is challenging who controls software discovery and execution more directly than it is replacing app-store billing./>
The distinction matters. The enterprise marketplace initially uses partner contracts and invoices, with no self-service checkout. Portable AI allowance also doesn’t magically pay for a partner’s software subscription. Compute is not your entire business model. Sadly.
Your next user might not have eyeballs
OpenAI’s new Dots, powered by GPT-6 Astra, have cloud computers and browsers and can connect through its plugin ecosystem to more than 4,000 apps. That figure describes integrations—not 4,000 newly built interactive ChatGPT applications.
Still, this is the part that gets my developer brain buzzing: your product may need to serve someone clicking a panel and an agent invoking an operation.
Those are different design problems.
A beautiful interface can tolerate a little ambiguity. An agent-facing operation that sends invoices cannot.
I’d prioritize three things:
1. Separate reads from writes. Looking up a customer and emailing that customer should never be one fuzzy tool.
2. Make retries safe. Duplicate-action protection beats apologizing for three identical invoices.
3. Keep permissions explicit. Authentication is not authorization.
That last point is documented, not theoretical. Sign in with ChatGPT shares name, email, and a profile picture if available by default—not conversations, memory, files, or billing details. Data access and allowance sharing require additional authorization.
Your login callback is not a permission slip to rummage through someone’s CRM.
Hot Take: the app store isn’t the real target
The more interesting target is the moment before an app gets chosen.
If ChatGPT owns that moment, Apple and Google aren’t the only companies facing pressure. Every SaaS business whose moat includes habit, navigation, or a recognizable dashboard should pay attention.
Better task completion could make software brands less visible. The customer remembers the assistant; your service becomes a replaceable tool underneath it.
And developers aren’t escaping gatekeepers. They’re meeting another one, with review rules and recommendation power. OpenAI’s planned retirement of custom GPTs in favor of plugins also makes platform churn a concrete engineering risk.
Build the integration. I would. But keep business logic on your backend, measure repeat usage and reliable task completion, and resist confusing platform reach with revenue.
The exciting opportunity is fewer steps between intent and outcome. The dangerous bargain is letting someone else decide whether your product belongs in those steps.
