Development & Automation

Build an AI Plugin Once, Run It Anywhere: OpenAI, AWS, and Cursor Back One Standard.

OpenAI, AWS, Cursor, GitHub, and Vercel have released Agent Plugins 1.0, an open standard that bundles Agent Skills and MCP into one package that runs across compatible clients.

· Huuush

Build an AI Plugin Once, Run It Anywhere: OpenAI, AWS, and Cursor Back One Standard

Key takeaways

  1. 1OpenAI, together with AWS, Cursor, GitHub, Microsoft, and Vercel, announced Agent Plugins open standard 1.0.
  2. 2A plugin bundles a plugin.json manifest, a skills folder, and mcp.json in a fixed structure, so Agent Skills and MCP settings work as-is across compatible clients.
  3. 3The standard defines only packaging and discovery. Security checks such as installation, permissions, and signing are left to each client, and it is still a Working Draft.

The AI agent ecosystem has reached an important moment: deciding whose standard everyone will use. OpenAI, together with AWS, Cursor, GitHub, Microsoft, and Vercel, has announced Agent Plugins open standard 1.0. The idea is simple: "build once, run anywhere."

Illustration of an AI agent and the plugin standard
Agent Plugins: an open standard that lets one package work across multiple agent clients (source: The Next Web)
Loading X post.

Unable to load the X post. Please try again.

OpenAI Developers: official announcement of Agent Plugins 1.0 (X embed)

How One Plugin Connects ChatGPT and Cursor

A plugin is a folder with a single plugin.json manifest at its root. Inside, a skills/ folder (SKILL.md, scripts, reference material) and mcp.json (MCP server settings) sit in fixed locations. Compatible clients scan this predictable structure to discover and load the plugin.

Skills and MCP Stay the Same, Packaging Becomes One

Agent Plugins are a packaging layer that wraps both rather than replacing them.

  • Agent Skills: reusable instructions and procedures (the manual)
  • MCP servers: live connections to tools and data (the USB port)
  • Agent Plugins: the box that bundles both for distribution (the container)

In other words, a skill's instructions and MCP settings are bundled in a shared format so they keep working when moved to another client.

Why OpenAI, AWS, and Cursor Chose the Same Spec

It was announced by OpenAI, but it is really a multi-vendor coalition. Vercel led the proposal, and the technical steering committee started with five members: Amazon, Cursor, Microsoft, OpenAI, and Vercel. Google also joined as a Core Maintainer. Compatible clients on launch day are ChatGPT, Codex, Cursor, GitHub Copilot, AWS Kiro, and VS Code.

When One Extension Reaches Many Agents

1. Development costs drop. Until now, agent extensions had to be rebuilt for each client. With a standard, you build once and use it in all the major agents.

2. The barrier to entry in the agent market falls. A small developer can reach ChatGPT, Cursor, Copilot, and others at the same time with one piece of work.

3. But security is still on each client. Version 1.0 standardizes only how plugins are packaged and discovered. Installation, permissions, sandboxing, signing, and trust verification are left to each client. Since a fake skill has already slipped past a security scanner, verifying plugin safety is still an unsolved problem.

Between Cheers and Doubt: What 1.0 Still Leaves Open

Reactions are mixed, from welcoming ("we really needed this") to skeptical ("it's a thin standard, and client-specific extensions will matter most in the end"). The standard is still a Working Draft (1.0.0), and the manifest schema, packaging CLI, registry, and versioning are expected to be refined further.

After the Feature Race Comes the Standards Race

This signals that the agent ecosystem is moving from a feature race to a standards race. Once a standard takes hold, moving between agents gets easier and platform lock-in weakens. But security and trust issues remain, so it will likely take time to mature.

Related: Vercel announcement · The Next Web · AWS blog

Frequently asked questions

What is the Agent Plugins standard?
It is an open standard that packages Agent Skills and MCP settings in a shared format, so an agent extension built once can run on multiple compatible clients.
Which clients can use it?
At launch, ChatGPT, Codex, Cursor, GitHub Copilot, AWS Kiro, and VS Code are compatible.
Does the standard solve security problems?
No. Version 1.0 standardizes only packaging and discovery. Installation, permissions, sandboxing, signing, and trust verification are left to each client, so verifying plugin safety remains an unsolved problem.
Do Agent Plugins replace skills or MCP?
No. They are a packaging layer that wraps both, bundling a skill's instructions and MCP settings for distribution.

Related articles