Amazon, Cursor, Microsoft, OpenAI, and Vercel have joined forces around Agent Plugins, an open standard designed to make AI agent extensions easier to package and reuse across platforms.
The effort targets a familiar problem for developers building agent extensions: each product has had its own format, folder structure, and setup process. Agent Plugins proposes one shared package format instead.
What Agent Plugins Is Trying To Fix
AI agent extensions can be useful only if they can be discovered, installed, and reused in a predictable way. Until now, the source article says developers have had to adapt their work separately for different products because each one relied on its own conventions.
That fragmentation creates extra work before an extension can move from one environment to another. A developer may have reusable instructions, workflows, tools, or data connections, but still need to repackage them for each platform.
Agent Plugins addresses that packaging layer. It defines a single way to bundle an extension so that participating platforms can recognize the same basic structure.
The Package Format Starts With plugin.json
The standard is built around a directory that includes a manifest file named plugin.json. That manifest is the center of the package format because it gives the extension a shared structure for discovery.
Version 1.0.0 supports two components:
- Agent Skills, which handle reusable instructions and workflows.
- MCP servers, which connect agents to tools and data.
Those two components show the practical scope of the first version. One side covers what an agent should know how to do in repeatable workflows. The other side covers how the agent reaches outside itself to work with tools and data.
For developers, the important change is not a new capability in isolation. It is the possibility of packaging those pieces once in a form that can be reused across products that support the standard.
What The Standard Does Not Cover
The source article is clear that Agent Plugins is limited in scope. It covers packaging and discoverability, not every part of how an extension will operate once installed.
Specifically, the standard does not cover marketplaces, permissions, or runtime environments. That means the shared format does not settle how plugins are distributed, how access is approved, or exactly how each platform executes them.
This boundary matters. A common package format can reduce duplication, but platforms can still make their own decisions about how users find plugins, what controls apply, and how the runtime behaves.
The specification is being developed in the open on GitHub. That open development process gives the standard a public venue, while the actual adoption and implementation details remain tied to the companies and platforms involved.
Why The Company List Matters
The group behind Agent Plugins includes Amazon, Cursor, Microsoft, OpenAI, and Vercel. That mix matters because the stated purpose of the standard is reuse across platforms, and a shared format is more meaningful when multiple major AI and developer-tool companies participate.
The effort also highlights a notable absence: Anthropic. The source article points out that Anthropic created both the Model Context Protocol and Agent Skills as open standards and recently added its own plugin system to Cowork, its desktop tool for agentic knowledge work.
Despite that connection to the same underlying ideas, Anthropic is not part of the Agent Plugins effort, according to the source. That absence does not change what Version 1.0.0 defines, but it is relevant because Agent Plugins includes MCP servers and Agent Skills as supported components.
The Near-Term Meaning For AI Agent Extensions
Agent Plugins is best understood as a packaging standard, not a full operating model for agent software. Its value is in giving developers a common directory and manifest structure for extensions that include reusable workflows and connections to tools or data.
If the standard is adopted by the platforms backing it, developers would have less need to rebuild the same extension packaging for each product. The source article does not claim that marketplaces, permissions, or runtimes are unified, so those areas remain separate from the standard itself.
For now, the clearest takeaway is simple: Amazon, Cursor, Microsoft, OpenAI, and Vercel want AI agent plugins to become more portable at the package level. Agent Plugins is the proposed format for getting there.