On August 6, 2026, OpenAI, AWS, Cursor, GitHub, Microsoft, and Vercel launched something that should have existed years ago: Agent Plugins, an open standard for building AI agent extensions that work everywhere.
The pitch is simple: build a plugin once, deploy it across ChatGPT, GitHub Copilot, Cursor, VS Code, Kiro, and any other compatible agent client. No rewrites. No per-tool setup instructions. One package, multiple platforms.
This is the end of vendor lock-in for agent tooling.
What Are Agent Plugins?
Agent Plugins 1.0.0 is an open, vendor-neutral specification for packaging reusable AI agent extensions into one portable format.
A plugin is just a folder containing three things: - plugin.json — metadata (name, version, description) - skills/ — Agent Skills (reusable instructions for your agent) - mcp.json — MCP server configurations (connections to APIs, databases, tools) That's it. The format is intentionally minimal. Vercel initiated the proposal. AWS, Cursor, Microsoft, OpenAI, and Vercel then refined it together into a specification anyone can implement.
The initial steering committee has five core maintainers (Amazon, Cursor, Microsoft, OpenAI, Vercel), and no single company can control the roadmap. It's licensed openly, and contributions are public.
Why Agent Plugins Matter
Before this, every AI agent tool expected different folder layouts. GitHub Copilot wanted one structure. Cursor wanted another. ChatGPT had its own format. You'd build an extension for one client and start over for the next.
That fragmentation ended today.
Agent Plugins eliminates three problems:
1. Vendor lock-in. You're no longer trapped in one tool's ecosystem. Build once, use everywhere.
2. Wasted engineering time. No more rewriting the same plugin for six different clients. One package ships to all of them.
3. Ecosystem fragmentation. Developers can focus on building great skills and MCP integrations instead of managing platform-specific boilerplate.
This matters because AI agents are becoming the primary interface for developers. If you're building agent extensions—tools that agents need—you want them to be discoverable, portable, and usable across the entire ecosystem. Agent Plugins makes that possible.
How Agent Plugins Work
Agent Plugins bundle two things developers already use:
Agent Skills are reusable sets of instructions—what your agent should know how to do. A skill might be "connect to Slack" or "query a database" or "run a terminal command."
MCP (Model Context Protocol) servers are connections to external tools and data. They handle the plumbing: talking to APIs, managing authentication, fetching live data. MCP is maintained by Anthropic and has become the standard for agent-tool integration.
Agent Plugins packages these together. When a developer publishes a plugin, compatible clients (ChatGPT, Copilot, Cursor, etc.) can discover, install, and load it as one unit. The skills define behavior. The MCP servers provide the connections.
Compatible Tools at Launch
These platforms support Agent Plugins as of August 6, 2026: - Codex (OpenAI's agent CLI) - ChatGPT (web + desktop) - GitHub Copilot (VSCode, JetBrains) - Cursor (AI code editor) - VS Code (with agent support) - Kiro (Amazon's agent framework) Google announced support the same day. More platforms are coming.
Should You Build Agent Plugins?
If you're a developer or founder, ask yourself:
Do you have a tool, API, or service that AI agents should know about? (Yes → Build a plugin)
Are you frustrated by proprietary plugin formats? (Yes → Contribute to the standard)
Do you want your extension to work everywhere? (Yes → Agent Plugins solve this)
If you're evaluating agent tools for your team, check whether they support Agent Plugins. It signals the vendor's commitment to not locking you in.
Getting Started
The official specification and author guides are available at agent-plugins.org. The specification repository lives on GitHub, and contributions are welcome.
For implementation examples, start with tools like Cursor or GitHub Copilot, which have published their integration patterns.
This is day one of portable AI agents. The ecosystem will move fast from here.
For more on AI infrastructure and developer tools, read more on Bitroot.