OpenAI's app directory quietly became a plugin standard

OpenAI's ChatGPT App Directory turned one year old this month, more or less, and by the time it got there it wasn't really an app directory anymore. What launched in October 2025 as a proprietary way to get your product inside ChatGPT has, in three moves over ten months, folded itself into an open, cross-vendor packaging format that OpenAI helped write and doesn't fully control. That's an unusual arc for a platform holder with the leverage to just keep the walled garden.

The timeline

• October 6, 2025 — OpenAI previews the Apps SDK at DevDay, an MCP-based toolkit for building interactive apps that render and act inside ChatGPT, with Coursera, Canva, Zillow, Figma, and Spotify as launch partners.
• December 17, 2025 — OpenAI opens app submissions to any verified developer and launches the App Directory, a browsable surface at chatgpt.com/apps.
• July 9, 2026 — OpenAI migrates the App Directory into a "Plugin Directory" shared by ChatGPT and Codex. Existing app connections keep working; new listings arrive as plugins, which can bundle skills, MCP servers, and UI together instead of shipping as a single app entry.
• August 6, 2026 — Agent Plugins 1.0.0 ships: a vendor-neutral packaging spec for exactly that bundle (skills + MCP + optional UI), with OpenAI, Microsoft, Amazon, Cursor, and Vercel on the steering committee. ChatGPT and Codex are launch clients.

Read in order, it's a platform holder narrowing its own moat on a roughly nine-month cadence: proprietary SDK, proprietary directory, shared directory, shared spec.

What "app" even means now

The terminology is worth untangling, because OpenAI didn't retire "app" — it demoted it. In the new plugin architecture, a plugin is the thing you publish and the thing that shows up once in the directory for both ChatGPT and Codex. Inside that plugin, an "app" is specifically the piece that registers an MCP server connection (via an .app.json manifest) and optionally returns a UI resource for in-chat rendering. A plugin can also carry skills — instruction-and-resource bundles with no server behind them — or both an app and skills together.

So the Apps SDK didn't go away; it became the tool you'd reach for to build the app-shaped component of a plugin, rather than the whole unit of distribution. The directory doesn't list "apps" anymore. It lists plugins, and a plugin without any MCP server in it is a perfectly normal, publishable thing.

Why the timing isn't a coincidence

The July migration landed a month before the Agent Plugins spec did. That ordering matters: OpenAI restructured its own directory around a skills-plus-MCP bundle before the cross-vendor group it was co-drafting with had even published 1.0.0. Either OpenAI built its July format to the draft spec on purpose, or the two converged because there's really only one sane way to package "instructions plus a tool server plus optional UI" once you've decided to do it at all. Both readings point the same direction: the packaging shape was never OpenAI's alone to design, even before the joint announcement made that official.

It also means a developer targeting ChatGPT and Codex today isn't really building against "the Apps SDK" as a closed target. They're building a plugin that has to satisfy a manifest format five other companies signed off on — the same agent-plugins-spec that Cursor, GitHub Copilot, and VS Code are adopting for their own extension surfaces. Publish once, and in principle the same package is legible to more than one client family. That's the whole pitch of Agent Plugins as a standard, and OpenAI's own directory is now the proving ground for whether the pitch holds up in a product with real submission volume.

The parts that stayed OpenAI's

Not everything converged. The submission review process, the quality bar, featuring decisions, and the actual chatgpt.com/apps surface are still entirely OpenAI's call — the spec governs the package format, not who gets listed or promoted. And the compatibility picture is asymmetric in practice: Agent Plugins 1.0.0 is a Working Draft with no unified plugin store or cross-client security model yet, so "publish once" currently means "package once, then still submit separately to each directory that wants you." ChatGPT and Codex share a directory because they're both OpenAI products, not because the spec created a shared marketplace across vendors.

Worth naming plainly: Claude and Claude Code aren't on the Agent Plugins steering committee and weren't among its launch clients, even though the format packages two specs Anthropic originated — MCP and Agent Skills. That's the tension already on this site's tracker entry for the standard, and it's the same tension visible from the OpenAI side of the merger: a company adopting a shared format for its own convenience while the format's actual neutrality is still unproven.

What to build to, if you're building now

If you're shipping something into ChatGPT or Codex today, the practical guidance is simple: build to the Agent Plugins manifest, not to a ChatGPT-specific integration pattern. It's the format both of OpenAI's surfaces expect now, review moves through the plugin submission flow rather than a standalone app flow, and — if the spec's other adopters actually ship compatible clients — the same package has a shot at working somewhere else without a rewrite. The bet is unproven, but building to the older, single-client pattern is now the more expensive path, not the safer one.

For the wider state of that standard — governance, the missing store and security layer, who else has actually shipped a compatible client — see the Agent Plugins entry and the live tracker.