The Doorway That Decides

On October 6, 2025, OpenAI announced apps inside ChatGPT, moving third-party services into the conversation itself. According to Maxwell Zeff's reporting for TechCrunch, the launch includes services such as Spotify, Zillow, Canva and Coursera, alongside a preview of the Apps SDK for developers.

There is a useful product idea here: ask for something, bring the relevant tool into the exchange, and keep working without translating the request into another interface. There is also a distribution question. When the assistant supplies the service as well as the answer, who decides which service gets the introduction?

Two ways into the conversation

TechCrunch describes two distinct routes to an app. A user can name it explicitly—asking Figma to turn a sketch into a diagram, for example. Or ChatGPT can surface an app it considers relevant to the request, such as Spotify when someone asks for a party playlist.

Those are not equivalent acts of choice. Naming a service carries the user's preference into the conversation. Asking for an outcome leaves more of the selection to the platform. The interface may feel continuous; the decision about who participates has moved.

That distinction matters more than whether the result arrives in a chat bubble. In the Zillow demonstration described by TechCrunch, an apartment search produced an interactive map that users could explore through further conversation. This is more than a link pasted into an answer. It puts part of another service's working interface inside ChatGPT.

Credit where due: that could spare people real effort. A demonstration establishes the interaction being offered, however, not its reliability across messy searches or its effect on developers' businesses. A shorter route to a tool is a capability. Fair distribution is a separate claim requiring separate evidence.

Connection is not selection

OpenAI says the system uses the Model Context Protocol to connect external services, TechCrunch reports. That technical connection does not settle the editorial and commercial problem of choosing among them. Making a service callable and deciding when to call it are different jobs.

Zeff identifies the unresolved question directly: how will ChatGPT choose between competing services? His report does not establish that companies can buy preferential placement. It would be premature to describe the launch as a paid recommendation market. It would be equally premature to treat a helpful recommendation as proof that the selection process serves users impartially.

OpenAI's announcement does set out participation rules: published draft developer guidelines, plus design and functionality standards affecting prominence. Those are real criteria, not an absence of governance. They do not explain how competing eligible services will be selected in a particular conversation, or establish that the resulting distribution is impartial.

The privacy boundary deserves the same precision. OpenAI describes a first-connection prompt explaining potential data sharing, alongside requirements for data minimization and privacy policies. That is an announced disclosure mechanism, not merely a promise to think about permissions later.

TechCrunch nevertheless identifies uncertainty about how much conversation an app receives. That is not evidence that every app sees an entire chat. Nor does the existence of a connection prompt establish that users understand its consequences. The questions are what context each app receives in practice, how clearly the prompt explains it, and whether the protections work—not whether OpenAI announced any safeguards.

A coherent arrangement would make both decisions legible: why this service appeared, and what it receives. It would let people reject a suggestion or choose an alternative without turning a simple task into a negotiation with the interface. Developers, too, need to understand how the published criteria translate into visibility, rather than having to infer the rules from traffic.

The announced rules provide a starting point for those tests, not evidence that their practical outcomes are settled. The circle widens when more people can participate on terms they can understand and contest—not merely when more services fit inside one company's window.

The launch makes the doorway more useful. Whether it makes the people passing through it more independent remains unproven. For now, the old distribution problem has learned to speak conversationally.