# Vercel, OpenAI, Microsoft and AWS back a new agent plugin standard

By Steven Van · 2026-08-06

Agent Plugins 1.0 packages Agent Skills and MCP servers into one portable folder, already supported by ChatGPT, Cursor, Copilot and VS Code.

On August 6, 2026, [Vercel](<https://creatorstoolbox.com/tools/vercel>) announced Agent Plugins 1.0.0, an open, vendor-neutral standard for packaging extensions for AI agents. The idea is simple and overdue: instead of building a separate plugin for ChatGPT, another for Cursor, and another for VS Code, you build one folder that works everywhere. Vercel initiated the proposal, then refined it with representatives from Amazon Web Services, Anysphere, GitHub, Microsoft, and OpenAI. At launch, the format is already supported by ChatGPT, Codex, Cursor, GitHub Copilot, Kiro, and VS Code. For anyone who has built custom instructions or tool connections for a coding agent, this is the moment those investments stop being locked to a single app.

## One folder, one manifest, two building blocks

An Agent Plugin is just a directory. The only required file is a minimal `plugin.json` manifest that declares the plugin name and the spec version it targets, pointing at a JSON Schema hosted at agent-plugins.org. Around that manifest, the spec defines fixed locations for exactly two component types. Agent Skills, which are reusable instruction packages, live under a `skills/` directory, each skill in its own folder with a `SKILL.md` file plus optional scripts and reference material. MCP servers, which connect agents to external tools and services, are declared in an `mcp.json` file that supports stdio, Streamable HTTP, and legacy HTTP+SSE transports.

The clever part is the escape hatch. Clients that want to ship features beyond the shared baseline can add reverse-domain namespaced directories, something like `com.example.client/`, inside the same plugin. That client-specific content is ignored by everyone else, so a plugin can carry Cursor-only or VS Code-only extras without breaking portability for the rest.

## Why skills and MCP needed a shared package

Both component types already existed as separate specs. Agent Skills give an agent domain knowledge and repeatable workflows, things like a brand voice guide, a video export checklist, or a deployment runbook. MCP servers give it hands, letting it call APIs, query databases, or drive design tools. What was missing was a standard way to distribute the two together. Agent Plugins does not modify either spec; it defers entirely to the Agent Skills specification for skill format and to MCP for server behavior, and simply defines where each lives inside a portable package.

Version 1 is deliberately narrow. It covers discovery and loading of those two component types and nothing else. Installation flows, marketplaces, permission prompts, and policy controls all remain the client's job. That restraint is why six competing companies could agree on it: the standard is a floor for portability, not a replacement for anyone's product.

## Who is steering it, and the notable absence

Governance runs through a Technical Steering Committee with representatives from AWS, Cursor, Microsoft, OpenAI, and Vercel, with development happening in the open on GitHub. Independent coverage from Digital Applied notes that Google announced it was joining the committee on launch day. The name missing from the list is Anthropic. Claude Code popularized the plugin-as-folder pattern that Agent Plugins clearly echoes, and its own `.claude-plugin` format uses a similar but distinct layout. VS Code currently detects both formats, so in practice the ecosystems overlap, but the industry standard and the format that inspired it are, for now, parallel tracks.

## What builders should watch before installing everything

If you publish agent tooling, adopting the format is close to free: restructure your existing skills and MCP config into the standard layout and you instantly target every supporting client. Installing third-party plugins deserves more caution. The 1.0.0 spec defers signatures, provenance, and publisher verification to future work, and analysts have pointed out that a portable format also makes it cheaper to distribute a malicious plugin across many clients at once. Treat unknown plugins the way you would treat an unsigned dependency until trust infrastructure catches up.

For creators, the direction matters more than the plumbing. The agents inside your editor, your terminal, and ChatGPT are converging on a shared extension format, which means the ecosystem of ready-made skills and tool connections will grow faster and follow you between apps. The spec, schemas, and author guides are live at agent-plugins.org.

![Vercel](<https://www.google.com/s2/favicons?domain=vercel.com&sz=128>)

Vercel

The platform for frontend developers: deploy, preview, and scale web apps and AI agents with zero config.

[View Vercel →](<https://creatorstoolbox.com/tools/vercel>)

Sources

[Vercel Blog](<https://vercel.com/blog/introducing-agent-plugins>), [agent-plugins.org](<https://agent-plugins.org/>), [Digital Applied](<https://www.digitalapplied.com/blog/agent-plugins-1-0-open-standard-portable-ai-skills>)

[Read Vercel, OpenAI, Microsoft and AWS back a new agent plugin standard on Creators Toolbox](<https://creatorstoolbox.com/blog/vercel-agent-plugins>)

---
Canonical source: https://creatorstoolbox.com/blog/vercel-agent-plugins
