Why OpenAI, Cursor, Microsoft, and Vercel Are Teaming Up to Turn AI Plugins into the Next npm

AI coding tools have long suffered from fragmented plugin formats, forcing developers to re‑package the same capabilities for each client such as Cursor, Codex, GitHub Copilot, or Kiro. To address this, OpenAI, Cursor, Microsoft, Amazon, and Vercel have jointly released an open standard called Agent Plugins , aiming to provide a single packaging format that multiple AI coding environments can recognize and load.

Five companies unify the “packaging box”

Agent Plugins is not a proprietary feature but a publicly maintained specification. A technical steering committee of five core maintainers—representatives from Amazon, Cursor, Microsoft, OpenAI, and Vercel (led by Vercel’s Jonathan Hefner)—was formed. The project started in April, was renamed to Agent Plugins in July, and version 1.0.0 was published on July 24, with public exposure on August 6.

plugin.json , containing Skill and MCP

The specification defines a simple directory layout:

my-plugin/
├── plugin.json
├── skills/
│   └── summarize/
│       ├── SKILL.md
│       ├── scripts/
│       └── references/
├── mcp.json
└── com.example.client/
    └── hooks/

Three key components are: plugin.json: identity metadata (name, version, author, license, target spec version). skills/: one or more Agent Skills, each containing commands, scripts, and reference material. mcp.json: configuration for the MCP server the plugin needs to connect to.

A minimal plugin may consist of only a plugin.json and a SKILL.md. Clients can still add proprietary extensions (e.g., custom hooks) in separate namespaces without polluting the shared portion.

This feels a lot like npm

npm solved more than just JavaScript distribution; it gave packages a unified name, version, dependency list, and entry point, and a public registry for discovery and installation. Agent Plugins mirrors the first step: a common package format that multiple tools can consume.

Future scenarios include a database plugin that bundles an operation spec, SQL‑checking Skill, and a read‑only MCP server, or a Vercel plugin that bundles platform knowledge, deployment flow, and project diagnostics. The same source can be read by Cursor, Codex, Copilot, and Kiro, each extracting the parts they support.

Currently listed compatible clients are VS Code, Cursor, GitHub Copilot, ChatGPT, Codex, Kiro, Hermes Agent, and OpenClaw. Clients can gradually adopt Skill types or MCP transport methods without waiting for a full feature set.

The most subtle thing: Anthropic is missing

The author notes the absence of Anthropic from both the initial committee and the compatibility list, even though Claude Code would naturally fit the Skill/MCP model. The project’s open governance on GitHub means Anthropic can join later, but the current lineup signals that AI‑coding tool competition is shifting from models and editors toward plugin entry points.

Still far from a real npm

While the direction is promising, Agent Plugins 1.0 only standardizes package layout and retrieval. It does not yet define a plugin marketplace, install commands, dependency resolution, automatic updates, permission controls, or trusted signing—components that constitute the hardest parts of a plugin ecosystem.

A Skill can influence an Agent’s reasoning, and an MCP server can launch local processes, access networks, and read data, raising supply‑chain attack risks that could exceed those of the traditional npm ecosystem if a universal plugin registry emerges.

Thus, Agent Plugins today resembles a newly defined package.json without an accompanying registry or package manager. Nonetheless, the effort mirrors the multi‑year standardization journey of the front‑end ecosystem, and if a universal AI plugin registry appears, installing an AI workflow could become as effortless as running npm install today.