Agent Plugins aims to make AI skills portable

Vercel has published the initial draft of Agent Plugins 1.0, a packaging standard backed by Amazon, Cursor, Microsoft, and OpenAI that aims to make agent skills and tools portable across platforms. The proposal was posted Thursday and quickly adopted by the Linux Foundation’s Agentic AI Foundation (AAIF), although it will remain an independent project.

The goal is straightforward: developers should be able to write an agent capability once and move it between compatible clients instead of rebuilding it for every product. VS Code, Cursor, GitHub Copilot, ChatGPT, Codex, and Kiro currently support the standard, according to The Register.

The effort also extends OpenAI’s Agent Plugins standard, which our previous coverage identified as an open format for portable skills and MCP servers across compatible AI products.

Recommended reading

Claude Code sessions can now message each other

How Agent Plugins packages skills and tools

Agent Plugins combines two existing pieces of agent infrastructure:

• Agent Skills, originally created by Anthropic in 2025, provide the instructions for completing a task.
• Model Context Protocol (MCP) supplies connections to external resources such as databases, APIs, shell scripts, and other services.

Version 1.0 puts both into a single directory-based package. Each plugin gets a top-level directory containing a JSON manifest with its name, version, and schema. Task instructions go in a skills subdirectory, while MCP server settings live in an mcp.json file. Authors can use a namespace directory for additional extensions.

The structure follows the file-system-as-configuration approach associated with Next.js. Bundling the instructions and MCP configuration is intended to make a plugin easy to copy into any supporting agent client.

“This keeps the shared format predictable without requiring it to absorb every client-specific experiment happening across agent products.”

The format deliberately stays narrow. Experimental hooks, rules, and extensions can be placed in the namespace directory, but their portability is not guaranteed. The AAIF may consider successful extensions for later versions, without promising that they will be added.

Portability still depends on the client

Agent Plugins standardizes the package, not the entire user experience or execution environment. Each agent platform still has to provide its own interface, security boundaries, and plugin marketplace. Supporting a common directory format therefore does not by itself ensure that a plugin behaves identically in ChatGPT, Cursor, or VS Code.

The project will use open governance, with an independent technical oversight committee considering public input. Engineers from Amazon, Cursor, Microsoft, OpenAI, and Vercel currently form the core development team, and the specification is licensed under CC-BY-4.0.

Agent Plugins is also entering a crowded standards effort. Google’s Agent2Agent (A2A) communications work has moved into the AAIF, while Google is developing Agentic Resource Discovery for finding tools on the web. The foundation is separately promoting Skills Over MCP, which embeds skills inside an MCP server rather than attaching an MCP configuration to each skill package.

The reporting does not establish whether the supported clients enforce the same security policies or deliver consistent behavior for the same plugin. That is the key unresolved limitation: Agent Plugins makes distribution materially cleaner, but its “write once, run everywhere” promise still depends on platform-level implementation outside the specification.