Build a plugin once and use it across compatible agent clients. Introducing Agen

Agent Plugins is a shared package format meant to let one add-on work in several AI agent apps without being repackaged for each. OpenAI announced it on August 6 as an open standard developed with AWS, Cursor, GitHub, VS Code and Vercel. It packages two existing kinds of components: Agent Skills, which are reusable instructions and resources, and configurations for MCP servers, which connect agents to tools and services.

The problem it addresses is duplicated packaging work. According to the standard's site, agent apps had each developed their own plugin formats even when the contents were the same, so authors had to rearrange or copy components for each one. Version 1.0.0 defines a directory with a required plugin.json manifest naming the plugin and the specification version, a skills folder, and an mcp.json file describing servers. A compatible app checks for the manifest and loads whichever component types it supports.

The standard is small on purpose. Commands, hooks and custom agents remain specific to each app, as do installation, permissions and user experience; an app can add its own features in a separately named folder that other apps ignore. A plugin built only from skills and MCP servers should therefore travel between supporting apps, while one that leans on a single app's extras will not. At launch, Vercel, which initiated the proposal, listed support in ChatGPT and Codex, Cursor, GitHub Copilot, Kiro and VS Code.