Five AI rivals agreed on one plugin standard, minus Anthropic

Six logos, one folder layout

On August 6, @OpenAIDevs posted a spec release with a very short pitch: "Build a plugin once and use it across compatible agent clients." The standard is called Agent Plugins. The list of names attached to it is the actual news. AWS, Cursor, GitHub, VS Code and Vercel all worked on the thing, and @vercel announced it the same minute as "an open standard for extending agents."

Google put out its own post that day, joining as a core maintainer with two products already shipping the format.

Then Tibo Sottiaux, who leads Codex at OpenAI, wrote a line worth reading twice: "A standard for (most) of your agents out there." The parenthetical is carrying more weight than it looks.

What a plugin means in this context

An AI agent on its own knows language and not much else. To do anything useful at your company it needs two extra things: instructions for how you do a specific job, and connections to the software where your data lives.

The industry already had a format for each. Agent Skills is a folder with a SKILL.md file that tells an agent how to handle a task, loaded only when that task comes up. MCP is the connector protocol that lets an agent read and write in your calendar, your CRM, your invoicing system.

What did not exist was a box to put both in. If you wrote a skill plus a connector for handling permit paperwork, you repackaged it separately for every tool that might run it.

Agent Plugins 1.0 is that box, and it is deliberately dull. A folder, a plugin.json manifest requiring exactly two fields, a skills/ directory, an optional mcp.json for connectors. Nothing clever. Clever is the enemy of a standard people actually implement.

The name that isn't on it

Agent Skills, the format this whole spec defers to for the interesting half, came out of Anthropic. The company published it as an open standard in December and handed the specification to a public site.

Anthropic is not on the technical steering committee, which is AWS, Cursor, Microsoft, OpenAI and Vercel. Claude Code is not in the launch client list either, which reads ChatGPT, Codex, Cursor, GitHub Copilot, Kiro and VS Code.

My read is less dramatic than the framing invites. This is roughly what success looks like when you open a format. Anthropic gave the thing away, its four largest competitors built the distribution layer on top of it, and the format won on the merits while the company that wrote it lost the steering wheel. That trade has been made before. Google published Kubernetes and then spent years as one vote among many. The trade is usually worth it, because a format nobody else adopts is just a folder convention with good documentation.

The part I would actually flag is further down the spec. Permissions, trust policy and sandboxing are explicitly out of scope. The text says the containment rules do "not sandbox a plugin subprocess." So a plugin is now portable across six clients, and what that plugin is allowed to touch is decided separately by each one. Portable code, non-portable safety. That gap gets filled by version 1.1 or by an incident, and standards bodies do not usually get to pick which.

What this means if you run a business

Very little this month. Quite a bit over the next year, in a specific way.

Your vendors are about to start shipping plugins. Your scheduling software, your accounting tool, your field service app will publish a package that any of these agents can install, the same way they all eventually shipped a mobile app. When that happens, the thing you were quoted a custom integration for in 2025 becomes a one line install.

Two consequences worth planning around. First, you get less lock-in, because a skill written for one agent now moves to another. Second, install becomes a real decision with real blast radius, because the spec just told you it is not sandboxing anything.

Here is the part no standard solves. A plugin can only encode a process that exists. If the way your shop handles a callback, a quote revision, or a past due invoice lives in one person's judgment and nowhere else, there is nothing to package. The companies that get value from this are the ones that already wrote down how the work goes.

That is most of what a free process audit at New Face Design turns up: not which AI tool to buy, but which of your repeated jobs is actually described well enough to hand to software. Do that part now and the install button will be worth pressing when it shows up.