Key Takeaways
- Agent Plugins 1.0.0 standardizes MCP server packaging into portable plugins.
- Claude Code users can create a plugin.json manifest to share tools across ChatGPT, Cursor, Copilot, and more.
What Changed
Vercel just shipped Agent Plugins 1.0.0 — an open, vendor-neutral standard for packaging Agent Skills and MCP servers into portable plugins. Think of it as the .zip for AI agent tooling: a single package that any compatible agent client can discover and load.
The spec defines a root plugin.json manifest plus fixed directory locations for portable components. Each client (Claude Code, Cursor, Copilot, etc.) keeps control over installation, distribution, policy, and UX — but the plugin itself is fully transportable.
Launch clients: ChatGPT, Codex, Cursor, GitHub Copilot, Kiro, and VS Code. Notably, Claude Code is not on the list — yet. But that doesn't mean you should ignore this.
What It Means For You
If you build MCP servers or Agent Skills (custom slash commands, tools, or workflows for Claude Code), you've probably hit this pain: you write a great MCP server, but sharing it with a teammate who uses Cursor or Copilot means duplicating config, docs, and setup steps.
Agent Plugins changes that. You package your MCP server once — with a plugin.json manifest and the server code — and any compatible client can load it. For Claude Code users, this is a future-proofing move: the standard is open, and Anthropic's own MCP ecosystem (which Claude Code heavily relies on) is a natural fit. When Claude Code adds support (likely soon, given MCP's ubiquity), your plugins will just work.
Even today, you can use the spec to structure your MCP servers more cleanly. The fixed layout forces you to separate manifest, server code, and assets — which is good practice anyway.
Try It Now
1. Read the spec at agent-plugins.org — it's short and implementation-focused.
2. Convert an existing MCP server. Suppose you have a Claude Code MCP server in ~/mcp/my-tool/. Create a plugin.json:
{
"name": "my-tool",
"version": "1.0.0",
"description": "Does X, Y, Z",
"components": {
"mcpServers": {
"my-tool": {
"command": "node",
"args": ["server.js"]
}
}
}
}
Then structure your repo:
my-plugin/
├── plugin.json
└── mcp/
└── server.js
3. Test with a supported client like Cursor or VS Code (if you use them) to verify the plugin loads. If you don't, at least validate the manifest against the spec.
4. Publish and share — drop it on GitHub, npm, or your internal registry. Anyone with a compatible client can install it without reading your docs.
Why This Matters for Claude Code Users
The agent tooling ecosystem is consolidating. MCP became the standard for connecting models to tools; Agent Plugins is becoming the standard for packaging those connections. Claude Code's strength is its deep MCP integration — 86 of our articles reference Claude Code using MCP. By adopting Agent Plugins early, you ensure your tools are portable when the standard lands in Claude Code.
Also note: the MCP ecosystem is shifting toward minimalism (fewer, better servers). Agent Plugins encourages that by making plugins lightweight and composable.
Source: vercel.com
[Updated 08 Aug via vercel_blog]
The standard is backed by a multi-vendor Technical Steering Committee including core maintainers from AWS, Cursor, Microsoft, OpenAI, and Vercel, with contributions from Anysphere and GitHub. The spec defines a minimal manifest requiring only name and specVersion, while component validation is independent—one invalid MCP server won't disable other skills. Client-specific extensions use namespaced namespaces, keeping portable parts clean. The GitHub repository at agentplugins/agent-plugins-spec hosts governance and contribution processes, reinforcing the open, vendor-neutral claim. [per Vercel]
