Agent Plugins 1.0: Build Once, Run in Any AI Agent

Agent Plugins 1.0, a packaging standard for AI agent plugins agreed upon by Amazon, Cursor, Google, Microsoft, OpenAI, and Vercel, was released on August 6 and became generally available across GitHub's Copilot ecosystem on August 12. The standard defines a single directory format with plugin.json, skills/, and mcp.json, allowing developers to build once and run in any compatible AI agent client, reducing the need to maintain multiple plugin formats. It does not replace MCP but complements it by providing a container for tools and skills.

Six of the biggest names in AI tooling — Amazon, Cursor, Google, Microsoft, OpenAI, and Vercel — agreed on a packaging standard for AI agent plugins while the official standards bodies were still scheduling their first meeting. Agent Plugins 1.0 https://agent-plugins.org/specification dropped August 6, went GA across GitHub’s entire Copilot ecosystem on August 12, and it quietly solves a problem every developer building AI integrations has been fighting: maintaining the same plugin in five different formats for five different clients. The Problem: Your Plugin, Packaged Five Times If you’ve shipped a skill or tool integration for more than one AI agent client, you already know what Agent Plugins 1.0 fixes. Before this spec, packaging the same capability for Cursor, GitHub Copilot, and ChatGPT meant separate directory layouts, separate manifest schemas, and separately shaped MCP configurations for each platform. Same code. Three to five copies of the wrapper around it. The maintenance burden only compounds as more agent clients launch. Agent Plugins collapses that into one directory that any compatible client can install. What the Standard Actually Is Agent Plugins 1.0 is a packaging format, not a new protocol. A plugin is a directory with three things: plugin.json — the portable manifest required skills/ — Agent Skills, one sub-directory per skill optional mcp.json — MCP server declarations optional A working plugin.json needs four fields: { "$schema": "https://agent-plugins.org/schemas/1.0.0/plugin.schema.json", "name": "my-dev-tools", "description": "Deployment utilities for React apps", "version": "1.0.0" } That’s the whole manifest. The deliberate minimalism — distribution, security, and UX left entirely to each platform — is precisely what allowed six vendors to agree in weeks rather than years. This Isn’t Replacing MCP The most common misread is that Agent Plugins 1.0 competes with MCP. It doesn’t. MCP defines how agents and tools communicate. Agent Plugins defines how you ship those tools — alongside skills — in one installable package. An Agent Plugin carries an MCP server in its mcp.json . Think npm package vs. the code inside it: one is the container, one is the protocol. Where It Works Today As of August 12, Agent Plugins 1.0 is generally available in: - GitHub Copilot in VS Code - GitHub Copilot CLI copilot plugin install <name - GitHub Copilot app - GitHub Copilot SDK The broader compatibility story — per the GitHub changelog https://github.blog/changelog/2026-08-12-agent-plugins-1-0-in-vs-code-copilot-cli-and-the-copilot-app/ and founding member announcements — extends to ChatGPT, Codex, Cursor, Kiro, and Vercel’s AI integrations. If a client has adopted the spec, your one plugin directory works there too. Migrating Existing Plugins If you maintain a GitHub Copilot plugin that predates Agent Plugins 1.0, nothing breaks. Backward compatibility is preserved. The VS Code documentation https://code.visualstudio.com/docs/agent-customization/agent-plugins describes adoption as “mostly manifest work”: add a conforming plugin.json at the root and you’re in. The Vercel announcement https://vercel.com/blog/introducing-agent-plugins confirms that existing plugins outside the spec remain fully supported with no forced migration timeline. What 1.0 Leaves Out Intentionally Version 1.0 ships with no central registry, no signature verification, no permission controls, no sandboxing, and no publisher identity checks. These are listed as future work on the spec roadmap, not oversights. Excluding them is what made cross-vendor consensus possible in the first place. The practical implication: treat Agent Plugin installs like any third-party dependency. Audit before you deploy, especially in enterprise environments. Platforms are expected to add their own trust layers; the standard deliberately leaves room for that. The Verdict Agent Plugins 1.0 is worth adopting now. The standard is minimal enough to implement in an afternoon, cross-client support already covers every major AI coding tool, and the cost of staying fragmented only grows as more agent clients enter the market. If you’re actively maintaining plugins across multiple clients, the consolidation is immediate. If you’re just starting out, building to the spec from day one costs nothing extra and future-proofs your distribution. The spec lives at agent-plugins.org https://agent-plugins.org . The example plugin and migration guide are on GitHub at agentplugins/agent-plugins-example .