Agent Plugins: An Open Standard for Skills and MCP

OpenAI Developers has introduced Agent Plugins, an open standard intended to let a plugin be packaged once and used across compatible agent clients. The announcement says the format was developed with Amazon Web Services, Cursor, GitHub, Code, and Vercel, and that it packages Agent Skills plus MCP server configurations in a shared format. (OpenAI Developers on X)

The project’s durable Agent Plugins specification page describes the format as open and vendor-neutral. Version 1.0.0 establishes a common package structure for Agent Skills and MCP servers, so clients can discover and load those components consistently instead of requiring authors to rearrange the same files for every host.

What the standard is trying to solve

Agent clients have developed different ways to package extensions, even when those extensions contain the same underlying skills or MCP connections. That forces authors to duplicate components or adapt a package before another client can use it.

Agent Plugins defines a small interoperability layer for the parts that can travel between clients. The standard does not attempt to make every client identical. Distribution, installation, permissions, user experience, and client-specific capabilities remain under the control of each client. (Agent Plugins)

That boundary matters. A portable package can reduce format work for authors, but it does not mean that installing a plugin in one client automatically grants the same permissions or produces the same interface in another.

What goes in a plugin

The Agent Plugins page describes a plugin as a directory with a required manifest and optional components in fixed locations:

my-plugin/
├── plugin.json
├── skills/
│   └── summarize/
│       ├── SKILL.md
│       ├── scripts/
│       └── references/
├── mcp.json
└── com.example.client/
    └── hooks/

The pieces have distinct jobs:

• plugin.json identifies the plugin and the Agent Plugins version it targets.
• skills/ contains Agent Skills using the format defined by the Agent Skills specification.
• mcp.json describes MCP servers using stdio, Streamable HTTP, or legacy HTTP+SSE transports.
• Reverse-domain extension namespaces give an individual client room to add behavior without changing the portable core.

This structure gives authors a predictable place for reusable skills, MCP configuration, and client-specific additions. It also leaves room for a client to ignore an optional extension when it does not support that capability. (Agent Plugins)

What developers can do with it

For a plugin author, the immediate appeal is less packaging duplication. A single directory can carry a skill and its MCP configuration in the shared layout, while compatible clients decide how to expose, install, and authorize it.

For a platform or internal-tools team, the standard offers a clearer contract for distributing reusable agent components. Teams can separate the portable package from host-specific policy instead of baking permissions and interface assumptions into every copy.

The standard is also relevant to MCP deployments. Because mcp.json can describe stdio, Streamable HTTP, and legacy HTTP+SSE servers, a plugin can point a compatible client at different kinds of MCP server configuration without changing the top-level package shape. The page describes the format; it does not promise that every client supports every transport or extension.

What Agent Plugins does not promise

Agent Plugins is a packaging and interoperability standard, not a universal marketplace or a cross-client permission system. The official page explicitly keeps distribution and installation under individual client control. It also leaves user experience and client-specific capabilities to each host. (Agent Plugins)

That means developers should still check a client’s compatibility list, supported transports, authentication behavior, and permission model before shipping a plugin. The OpenAI Developers announcement names organizations involved in developing the standard, but it does not by itself establish that every named organization supports every part of the format in every product. (OpenAI Developers on X)

Open governance

The specification page says Agent Plugins is openly licensed and developed in public. Its initial Technical Steering Committee includes core maintainers from Amazon, Cursor, Microsoft, OpenAI, and Vercel. Proposals and technical decisions are public, and participation is open to the broader ecosystem. (Agent Plugins)

That governance model is designed to keep the portable core separate from vendor-specific extensions. New features and material changes begin in public discussions, where a proposal can establish a concrete portability need and implementer support.

Bottom line

Agent Plugins gives Agent Skills and MCP server configurations a shared package format instead of asking authors to maintain a separate layout for every agent client. Its version 1.0.0 core is intentionally narrow: a manifest, predictable component locations, and a path for client extensions. The practical benefit will depend on which clients adopt the format and how they handle installation, permissions, transports, and host-specific features.

Sources

• OpenAI Developers, “Introducing Agent Plugins,” August 6, 2026: https://x.com/OpenAIDevs/status/2085398373511918022
• Agent Plugins official specification page: https://agent-plugins.org/