OpenAI updates plugin creation and discovery, but a public listing still needs review

OpenAI's [September 29 DevDay recap](https://openai.com/index/devday‑2026‑recap/) says Plugin Creator will help developers build plugins, a redesigned submission flow will give clearer feedback, and improved ranking and recommendations will help people find relevant plugins. For a builder, those are three different milestones: a package that works in local testing, a submission that passes review, and a public plugin that a user actually discovers and chooses to install. Treating any one of them as an automatic path to the next would lead to a premature launch plan. The distinction matters most for teams with an existing MCP server. OpenAI's [plugin architecture guide](https://developers.openai.com/plugins/concepts/plugins) says a plugin can contain skills, an MCP server, or both, with an optional interface. The new creator is a packaging aid, not a substitute for designing a useful tool, testing its permissions or meeting directory requirements. A team should first decide what user task its plugin handles, then choose the smallest package that exposes that task reliably.

OpenAI's [packaging guide](https://developers.openai.com/plugins/build/plugins) recommends `@plugin‑creator` for the fastest OpenAI‑specific setup. It can scaffold a supported plugin manifest and a personal marketplace entry for testing. The same guide makes a boundary easy to miss: local and repository marketplaces are authoring or team‑distribution sources, while a public plugin goes through the universal Plugins Directory shared by ChatGPT and Codex. Installing a local entry proves the package can be tested; it does not prove that OpenAI reviewed it or that other users can find it. For an MCP‑backed plugin, the builder still needs a working server connection. The [submission reference](https://developers.openai.com/plugins/deploy/submission) accepts skills‑only, remote‑MCP‑only and combined submissions, and says a submitted remote MCP server must have a stable public HTTPS endpoint. That is a practical fork in the build plan: a skills‑only package and a plugin that reaches customer data through MCP have different integration and test work, even though both can appear in the same directory after approval.

The current [submission guide](https://developers.openai.com/plugins/deploy/submission), checked September 30, requires a verified developer or business identity for public submissions. A non‑owner submitter needs Apps Management write access in the owning Platform organization. The portal collects a public listing, availability and policy details; the guide asks for five positive and three negative test cases. For remote MCP, it also calls for accurate tool annotations, domain verification and reviewer access that can reproduce the tests. A polished manifest alone cannot clear those checks. Submitting starts OpenAI's review; it does not publish the plugin. The developer chooses when to publish after approval, according to the same [public publishing flow](https://developers.openai.com/plugins/deploy/submission). That gives a team a useful release boundary: finish the real user workflow and review materials before announcing availability, then verify the approved version and directory listing after the publisher acts.

The DevDay recap promises improved ranking and recommendations, including discovery in the directory and in conversations, while saying users choose which plugins to use and approve the access each receives. OpenAI's [plugin guidelines](https://developers.openai.com/plugins/plugin‑guidelines) describe stronger placement or proactive suggestions as possibilities for plugins with strong utility and satisfaction, not a benefit conferred by submission. Builders should therefore judge the launch by whether the plugin solves a specific user request and can be found for relevant prompts, rather than count on a placement that OpenAI has not promised. OpenAI's [metadata guide](https://developers.openai.com/plugins/guides/optimize‑metadata) recommends a labeled set of direct and indirect prompts, with expected calls and non‑calls, to test whether names and descriptions make a tool likely to be invoked for the right job without accidental activations. That is a more concrete next step than simply adding keywords to a listing. This announcement belongs with [ChatGPT](https://opentools.ai/tools/chatgpt) and its maker [OpenAI](https://opentools.ai/organizations/openai) in the OpenTools catalog; Plugin Creator and submission are parts of that platform workflow, not newly named foundation models or separate OpenAI tools.