Vercel-Led Standard Lets AI Agents Carry Skills Between Tools

Vercel and five other companies shipped Agent Plugins 1.0.0 on Aug. 6, a shared packaging format for AI agent skills and MCP server configs that lets a developer build a plugin once and install it into any compatible agent client, instead of wiring the same tooling into ChatGPT, Cursor, VS Code and Copilot separately (Vercel changelog, Aug. 6, 2026).

What the format actually does

An Agent Plugin is a directory with a plugin.json manifest (carrying just a name and a version) plus fixed locations for the plugin’s Agent Skills and MCP server definitions. A compatible client reads the manifest, discovers what’s inside, and loads it (Vercel blog, “Introducing Agent Plugins”). At launch it works in ChatGPT and Codex, Cursor, GitHub Copilot, Kiro and VS Code. The spec is deliberately narrow: it standardizes packaging only, leaving marketplace listings, permissions and discovery to each client to solve on its own (Vercel blog, “Introducing Agent Plugins”).

Google joined the same day as the sixth core maintainer, represented on the technical steering committee by Kevin Hou, a senior staff manager at Google DeepMind, alongside Amazon, Cursor, Microsoft, OpenAI and Vercel (Google Developers Blog, Aug. 6, 2026). AWS framed the process itself as the point: “This is how the best standards emerge. Not from a single vendor’s vision imposed on the ecosystem, but from the convergence of shared problems solved together” (AWS Open Source Blog, Aug. 6, 2026).

Vercel shipped a plugin of its own on launch day: running npx plugins add installs more than 47 Vercel-authored skills and subagents into Claude Code or Cursor, giving those tools direct access to Vercel’s deployment tooling. Vercel co-founder Guillermo Rauch called the result “production deployment superpowers” for those clients (Rauch on X, Aug. 6, 2026).

Whose idea was it

The companies don’t agree on the credit. Vercel’s own blog post says Vercel initiated the proposal and that AWS, Cursor, GitHub, Microsoft and OpenAI refined it from there (Vercel blog, “Introducing Agent Plugins”). OpenAI’s own announcement described it differently, framing Agent Plugins as something it built “with” the other five partners rather than crediting Vercel as the originator (OpenAI Developers on X, Aug. 6, 2026). Neither account changes what shipped, but it’s a reminder that a “joint” standard between six competitors still comes with six PR departments.

Our take

A shared packaging format is boring and useful, the kind of plumbing that’s easy to under-appreciate right up until every agent vendor has reinvented it slightly differently and nothing installs anywhere else. The narrow scope is the smart part: by refusing to standardize discovery, trust or permissions, the committee sidestepped the fights that kill standards (whose marketplace wins, whose review process is the review process) and shipped something six competitors could actually agree to. The tradeoff is that “compatible with Agent Plugins” will tell a developer almost nothing about whether a given plugin is safe to run, since none of that layer exists yet. Format compatibility isn’t safety, and this spec doesn’t pretend otherwise.

Key Takeaways

• Agent Plugins 1.0.0 launched Aug. 6, packaging Agent Skills and MCP server configs into one format installable across ChatGPT/Codex, Cursor, GitHub Copilot, Kiro and VS Code.
• The spec covers packaging only; discovery, trust and permissions are explicitly left to each client.
• The technical steering committee has six core maintainers: Amazon, Cursor, Google, Microsoft, OpenAI and Vercel.
• Vercel shipped a companion plugin with 47+ of its own skills for Claude Code and Cursor, installable via npx plugins add.