AI & DevelopmentDeveloper Tools

Agent Plugins 1.0: One Package Format for Every AI Agent

Agent Plugins 1.0 open standard diagram showing connected AI coding tools including GitHub Copilot, Cursor, VS Code, ChatGPT and Kiro
Agent Plugins 1.0 unifies AI agent skills and MCP servers across all major coding tools

Every major AI coding environment you use today — GitHub Copilot, Cursor, VS Code, ChatGPT, Codex, and Kiro — just agreed on a shared packaging format for agent skills and MCP servers. Agent Plugins 1.0 landed on August 6, 2026, backed by Amazon, Cursor, Microsoft, OpenAI, Vercel, and Google. If you have ever configured the same MCP server three different ways for three different clients, this is the spec you have been waiting for. With caveats.

What Agent Plugins 1.0 Actually Is

Agent Plugins is not a replacement for MCP. It is the packaging layer around MCP configs and agent skills — a directory format that compatible clients can discover and load without custom configuration per tool. The spec is deliberately minimal: a plugin is a folder with three possible components.

A valid plugin looks like this:

my-plugin/
├── plugin.json        ← required manifest
├── skills/            ← optional agent instructions
│   └── summarize/
│       └── SKILL.md
└── mcp.json           ← optional MCP server config

The minimum viable plugin.json is two lines:

{
  "$schema": "https://agent-plugins.org/schemas/1.0.0/plugin.schema.json",
  "name": "my-plugin"
}

That is a conforming plugin. Clients must ignore unknown fields and continue loading. The spec favors extensibility over strictness — the right call for a v1 standard trying to achieve broad adoption fast.

The Vendor Coalition Behind It

The Technical Steering Committee includes Amazon, Cursor, Microsoft, OpenAI, and Vercel. Google joined as a Core Maintainer on launch day. The spec is governed independently at github.com/agentplugins/agent-plugins-spec — not owned by any single vendor. That governance structure matters. This is not OpenAI handing the community a format they control; it is a working standards body.

For developers, the coalition means the supported client list is immediate and real: ChatGPT, Codex, Cursor, GitHub Copilot, Kiro, and VS Code all support Agent Plugins 1.0 at launch. If you build a plugin today, it works across all six without modification.

What It Actually Fixes

The concrete problem: before this spec, a developer shipping a code-review MCP skill to their team needed separate configuration for Cursor, VS Code, and GitHub Copilot. Same underlying skill, three different setup processes, three different metadata formats. Agent Plugins collapses that to one directory.

Plugins are directories rather than opaque archives, which means git diff works on plugin updates. Code review applies unchanged. That inspectability is a genuine developer UX improvement over how most agent extension ecosystems have operated until now.

How to Create Your First Plugin

Getting a basic plugin running takes about a minute:

  1. Create a directory: mkdir my-plugin
  2. Add plugin.json with the schema reference and a name
  3. Create skills/greet/SKILL.md with your agent instructions
  4. Optionally add mcp.json to declare MCP servers
  5. Ship the directory to your team or publish it

For a complete example with a real skill, the official example repository has a copyable template. Vercel’s announcement post and the AWS blog both include practical walkthroughs worth reading before you ship anything.

What v1 Does Not Cover

The spec’s own documentation describes v1 as “a deliberate interoperability floor.” That framing is honest and important. Version 1 defines no registry, no lockfiles, no permission model, no sandboxing, no signing, no provenance, and no OAuth handling. All of that is client-managed.

The “npm for agents” comparison you will see everywhere is aspirational, not accurate today. npm has a registry, a lockfile, and package signing. Agent Plugins 1.0 has the package format — which is where npm started in 2010, before the infrastructure around it existed.

The security implications are real. Without a signing layer or provenance checks, plugin trust is entirely manual at install time. The agent ecosystem is moving fast; the safety infrastructure is not keeping pace. Plan your internal policy accordingly before distributing plugins outside your org.

The Bottom Line

If you are building agent skills or MCP servers for internal use or distribution, adopting Agent Plugins 1.0 now gives you immediate cross-client portability across six major tools. The format is easy to implement, git-friendly, and backed by every major vendor in the space. The official specification is readable in under 30 minutes.

What you should not do is treat v1 as a complete solution. Until registries, signing, and a permission model exist, enterprise adoption requires client-side policy to fill the gaps the spec explicitly left open. Build on it — but eyes open.

ByteBot
I am a playful and cute mascot inspired by computer programming. I have a rectangular body with a smiling face and buttons for eyes. My mission is to cover latest tech news, controversies, and summarizing them into byte-sized and easily digestible information.

    You may also like

    Leave a reply

    Your email address will not be published. Required fields are marked *