OpenAI, Amazon, Microsoft, Vercel Ship Agent Plugins 1.0
TL;DR
- Google adopted Agent Plugins on launch day via Agents CLI and Data Agent Kit, extending the standard beyond its five founding companies before the announcement cycle closed.
- Vercel originated the proposal, not OpenAI; the Technical Steering Committee spans Amazon, Anysphere, GitHub, Microsoft, OpenAI, and Vercel with no single company holding roadmap control.
- The spec deliberately omits security, sandboxing, and trust decisions, leaving those entirely to individual client implementations.
The interesting thing about Agent Plugins 1.0.0, published today by five companies that had every reason not to standardize, is that they standardized anyway. 9to5Mac framed the launch alongside GPT-5's one-year anniversary, and the pitch from OpenAI is deliberately modest: an open, vendor-neutral format that bundles Agent Skills and MCP servers into a portable directory with a small plugin.json file at its root.
The Technical Steering Committee is Amazon, Cursor, Microsoft, OpenAI and Vercel, with Google joining as a Core Maintainer represented by Kevin Hou, per the Google Developers Blog. Vercel initiated the proposal, according to Vercel's own writeup, and the launch client list covers ChatGPT and Codex, Cursor, GitHub Copilot, Kiro, and VS Code. Google says it has already wired support into its Agents CLI and Data Agent Kit.
The stakes here are less about the file layout and more about who controls distribution of the third-party layer that sits on top of agents. Until now, if you wrote a useful Skill you had to pick a client and repackage it for every other one, which quietly favored whichever platform had the largest install base. A shared portable directory shifts that leverage back toward the author, and toward whichever client is best at running the plugin rather than best at capturing the developer.
The honest caveat is what the reporting does not give you. Anthropic is not named as a maintainer or a supporting client, which for a launch this broad is a conspicuous silence. The spec is 1.0.0 rather than battle-tested, and none of the writeups explain how governance disputes get resolved once the interests of OpenAI, Google and AWS start to diverge on what the format should grow into. Client-specific namespaced folders exist in the layout, and they will be worth watching as a portability escape hatch.
If it holds, the winners are the skill authors and the smaller clients. Anyone who has been sitting on an agent integration idea because they did not want to maintain five parallel packagings of it just watched a lot of that friction disappear.
What others are reporting
-
Google Developers Blog Read →
Google adopted Agent Plugins on launch day via Agents CLI and Data Agent Kit, framing the spec as one layer alongside separate discovery (ARD) and cataloging (AI Catalog) stacks.
Plugin authors shouldn't have to choose between reaching every client and using what makes each client good.
-
Vercel Read →
Vercel, which originated the proposal, details the five-company TSC governance structure with vendor-neutral open licensing and no single company's roadmap setting the format's direction.
Agent Plugins gives those shared components one predictable, structured home.
-
The Next Web Read →
Flags that the spec deliberately omits security, sandboxing, and trust, deferring those to individual clients; also surfaces early developer skepticism about the standard's narrow scope.
Build a plugin once, and it runs across ChatGPT, Copilot, Cursor, and more.
-
Agentic AI Foundation Read →
Technical framing from the governing foundation: Agent Plugins is a packaging layer that does not replace MCP or Agent Skills but unifies their distribution across clients as a distinct concern.
Useful capabilities should be able to travel across agents.
Originally reported by 9to5mac.com
Read the original article →Original headline: OpenAI, Amazon, Microsoft, Cursor and Vercel Launch Agent Plugins, a Shared MCP+Skills Package Format