OpenAI has introduced Agent Plugins, a shared format that packages Agent Skills and MCP server configurations for use across compatible agent clients. The company describes it as an “open standard” developed with AWS, Cursor, GitHub, VS Code and Vercel.
OpenAI introduces Agent Plugins standard for cross-client agent extensions
The format is intended to let developers build a plugin once and distribute it across multiple supported clients, rather than adapting the same extension separately for each platform. The accompanying material claims that Agent Plugins can package both agent capabilities and the configurations needed to connect MCP servers.
OpenAI’s launch list includes:
- Codex
- ChatGPT
- Cursor
- GitHub Copilot
- Kiro
- VS Code
The announcement does not list Claude Code or other Anthropic clients among the compatible platforms at launch. Several commenters questioned whether Anthropic had been invited to participate or might support the format later. The available announcement does not provide an answer.
That omission drew particular attention because commenters noted Anthropic’s role in creating and open-sourcing MCP. Other reactions welcomed a common packaging format, while some questioned whether another standard was necessary. One developer argued that a command-line interface and accompanying skills stored in the same repository could already cover much of the use case.
The initial client whitelist also prompted skepticism about how open the standard is at launch. One commenter characterized it as “a beta with extra steps,” arguing that a standard limited to named compatible clients has not yet demonstrated broad interoperability.
The announcement arrives with some familiar tension around technical standards. A separate graphic associated with the posts uses the standard-proliferation joke: a situation with 14 competing standards leads to a proposal for one universal standard, followed by a situation with 15. That framing reflects a concern raised in the discussion—that a format intended to reduce fragmentation could add another layer to an already crowded ecosystem.
For plugin developers, the proposed benefit is narrower and more concrete: a shared package for Agent Skills and MCP server settings could reduce duplicate implementation and maintenance work where clients support the format. Whether that benefit materializes will depend on adoption beyond the launch group, as well as on how the specification handles differences between individual clients.
OpenAI has not provided, in the announcement reproduced here, details about project-local installation, chat-specific scoping, translators for unsupported clients or broader adoption plans.
Source: OpenAI Developers on X
