Thought Leadership
Major AI vendors have agreed on a common open standard that allows extensions for AI agents to be used across different platforms and vendors for the first time.
On August 6, the project’s technical steering committee published version 1.0.0 of the specification. The proposal was originally initiated by Vercel and further developed with representatives from Amazon, Anysphere, GitHub, Microsoft, and OpenAI. The format combines two building blocks that are already widely used across the industry into a standardized folder structure: so-called Agent Skills, reusable instruction packages that provide AI agents with workflows and context for specific tasks, and MCP servers, which connect agents to external tools and data sources. Until now, developers had to package and adapt both components separately for each AI product, for example for Claude, Cursor, Gemini, or ChatGPT, even when the underlying functionality was identical.
At its core, an Agent Plugin consists of a folder containing a central manifest file called plugin.json, along with optional Skill and MCP configurations. ChatGPT, Codex, Cursor, GitHub Copilot, Kiro, and Visual Studio Code are among the platforms supporting the new format at launch. Google announced on the same day that it was joining the project’s steering committee and introduced two initial products based on the standard: Agents CLI and Data Agent Kit.
Deliberately Limited to an AI Packaging Format, With No Security Requirements
Notably, Anthropic is not among the developers of the new standard, even though both the Agent Skills concept and the underlying Model Context Protocol were originally developed by the company and released as open standards. According to industry reports, however, Agent Plugins can also be used with Claude Code because an accompanying command-line tool can automatically translate the portable format into Claude’s own plugin system.
The standard’s developers explicitly emphasize that version 1.0.0 is intentionally limited to defining a package format. The specification does not cover installation mechanisms, distribution channels, permission models, sandboxing requirements, or requirements for verifying provenance and trustworthiness. It also does not define how users should interact with the plugins.
This leaves it entirely up to companies, individual developers, and other organizations how they use the format in practice, particularly when it comes to security and permissions. Google’s own documentation also explicitly advises developers to consider whether they need a plugin at all before using one. If the goal is simply to provide a single MCP server for a single client, Google says a straightforward mcp.json file remains the simpler option.
(Editorial Team)
