Amazon, Cursor, Microsoft, OpenAI, and Vercel Launch Agent Plugins, a Shared Packaging Standard for AI Agent Extensions

Amazon, Cursor, Microsoft, OpenAI, and Vercel have jointly released Agent Plugins, an open standard for packaging and distributing AI agent extensions across platforms. The specification defines a single directory-and-manifest format that developers can use to bundle extensions once and reuse them across multiple agent products, rather than rebuilding for each platform's proprietary format.

What the standard defines

Agent Plugins version 1.0.0 is built around a plugin directory containing a manifest file called plugin.json. This version supports two component types:

• Agent Skills — reusable instructions and workflows that agents can invoke.
• MCP servers — connections to external tools and data sources via the Model Context Protocol.

The scope is deliberately narrow. According to the participating companies, the standard only addresses packaging and discoverability of extensions. It does not define marketplaces, permission models, or runtime execution environments — each platform retains control over how plugins are distributed, sandboxed, and run.

The specification is being developed in the open on GitHub, allowing outside contributors to propose changes and track its evolution.

The problem it addresses

Before Agent Plugins, every AI coding and agent platform used its own folder structures, manifest formats, and setup processes for extensions. A developer building a tool integration for one platform had to redo the packaging work to support another. By standardizing the package format, the five companies aim to let a single plugin work across their respective products without platform-specific rewrites.

Anthropic's absence

Anthropic is not part of the Agent Plugins coalition, despite having originated both foundational technologies the new standard builds on: it created the Model Context Protocol and introduced Agent Skills as open standards. Anthropic has also recently shipped its own plugin system for Cowork, its desktop tool for agentic knowledge work. The company's decision to build a separate system rather than join this effort was not explained in the announcement, and the reason for the omission remains unclear.

What this means

A shared packaging format could meaningfully reduce fragmentation in the fast-growing agent tooling ecosystem, where MCP servers and skill libraries have proliferated with inconsistent structures. If adopted broadly, Agent Plugins could let developers publish an extension once and have it work in Cursor, GitHub Copilot-adjacent Microsoft tools, Amazon's agent products, OpenAI's ecosystem, and Vercel's platform without modification.

But the standard's narrow scope is a limiting factor. Packaging is only one piece of the puzzle — marketplaces, permissions, and runtime behavior are arguably where the most consequential platform decisions get made, and those remain unstandardized. Two agents that both support Agent Plugins could still behave very differently once a plugin is installed.

Anthropic's absence is the more interesting signal. As the creator of MCP and Agent Skills, Anthropic sitting outside a coalition that builds directly on its own open standards suggests either a strategic disagreement over governance, a preference to control its own Cowork ecosystem, or simply timing. Whether Agent Plugins becomes a genuine industry standard likely hinges on whether Anthropic eventually aligns with it — without the company that originated the underlying protocols, the standard risks becoming one of several competing formats rather than the unifying one it aims to be.