Introducing Agent Plugins
vercel.com ● Covered by 2 sources
Vercel, OpenAI, Microsoft, AWS, GitHub and Anysphere just agreed on one packaging format for AI agent plugins. That means you build a Skill or MCP server once and it works in ChatGPT, Cursor, Copilot, Kiro and VS Code without repackaging.
Based on reporting by vercel.com — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
For the last year or so, building an extension for an AI agent has meant picking your battles. You write a Skill or wire up an MCP server, then you repackage it three or four different ways because ChatGPT wants its metadata one way, Cursor wants it another, and whatever client you forgot about wants something else entirely. Agent Plugins, which hit 1.0.0 today, is a direct answer to that annoyance, and it comes with an unusually broad list of backers for a spec this young.
The format itself is almost aggressively simple. A plugin is just a folder with a plugin.json manifest, a skills/ directory, and an mcp.json file, plus room for client-specific extras tucked into their own namespaced folders. The manifest needs exactly two fields to be valid: a schema reference and a name. Everything else is inferred from where files sit in the directory. That minimalism is the point. Jonathan Hefner and the rest of the team behind it, drawing on work from Vercel, AWS, Anysphere, GitHub, Microsoft, and OpenAI, deliberately kept the contract narrow so any client could implement it without much friction.
What's notable is what the spec refuses to touch. It doesn't try to reinvent Agent Skills or MCP, both of which already have their own specifications and real adoption. It doesn't govern installation, pricing, permissions, or UI. Commands, hooks, and other agent components are left out entirely for now, with the door open for the Technical Steering Committee to add them later if there's a real need. Clients also get a namespaced escape hatch for their own proprietary behavior, so nothing forces GitHub Copilot and Cursor to behave identically just because they share a packaging format.
The multi-vendor structure matters more than the technical details, honestly. Vercel proposed it, but the initial steering committee pulls Core Maintainers from AWS, Cursor, Microsoft, OpenAI, and Vercel, and the whole thing is openly licensed with a public governance process. At launch, support already spans ChatGPT and Codex, Cursor, GitHub Copilot, Kiro, and VS Code, which is a rare show of coordination among companies that are otherwise racing each other on every other layer of the agent stack.
Whether this actually sticks depends on whether extension authors bother adopting it before there's real pain from maintaining five separate plugin formats. Standards bodies have a long history of arriving right on time or five years too late. But with this many major clients committing on day one, Agent Plugins at least starts from a position most nascent specs would kill for.
My take — AI-written commentary, not fact-checked reporting
Every time five competing companies suddenly agree on a shared standard, it's worth asking what problem got expensive enough to force the truce, and here it's obvious: repackaging the same Skill for six clients is dumb, unpaid work nobody wants to keep doing. The restraint to leave MCP and Skills alone and just standardize the wrapper is the smart part, and it's the reason this might actually survive contact with reality instead of becoming another abandoned RFC. Still, a spec is only as good as its slowest adopter, and the real test is whether some client quietly starts shipping proprietary extensions in that namespaced escape hatch until it's not really portable anymore.
Read more about this at: vercel.com
