Packaging and discovery are settled; marketplaces, permissions and trust stay each client's problem.
OpenAI and four competing firms released Agent Plugins last week, a specification open to anyone under which a single agent extension runs inside rival AI products. From day one it works in OpenAI's ChatGPT and Codex apps and in Cursor, GitHub Copilot, Kiro and VS Code.
A plugin in the format bundles two pieces developers already use: Model Context Protocol servers, which hook an agent up to live data and tools, and Agent Skills, reusable instruction sets. That bundle travels further than the Codex plugins OpenAI shipped earlier, which function solely inside the company's own tools.
OpenAI stood at the centre of the public launch, but Agent Plugins is not an OpenAI product. The proposal began at Vercel, and the 1.0 specification was shaped by people sent from Amazon, Anysphere (the maker of Cursor), GitHub, Microsoft, OpenAI and Vercel. Five organisations sit on the body that steers it: Amazon, Cursor, Microsoft, OpenAI and Vercel. The licence is open, and those behind the project say no one firm's plans dictate where it goes.
The narrow scope is deliberate. The standard defines how a plugin is packaged and how clients discover it, and stops there, leaving storefronts, installation, permissions, sandboxing and trust to each client. That tight remit makes the format simple to take up and leaves the thorniest questions unanswered — among them whether a given plugin is safe to run. The concern is live: fake Agent Skills slipped past security scanners earlier this year.
Two developers landed on opposite sides of the format. Dax Raad, a builder of the developer-tools framework SST, pushed back hard, calling the standard shallow and arguing that whatever is useful in it will simply be pulled into client-specific extensions in the end. Angie Jones, a developer advocate, welcomed it, saying she had been looking for one way to carry her skills across the tools she works with.
Putting the underlying plumbing on a common footing leaves the competition question open. With one format in place, a lone developer might reach all the big agent platforms in a single step; equally, the few clients with the largest user bases may be dug in deeper, because whatever people are already running is what a standard favours.