OpenAI, Microsoft, and Cursor Unite Behind Agent Plugins to End Fragmented AI Workflows

Every major coding agent has had its own plugin format. A workflow packaged for GitHub Copilot needed rework before Cursor could use it. A tool integration wired up for Codex wouldn't load in VS Code. Plugin authors ended up maintaining parallel versions of the same thing across a fragmented ecosystem. Agent Plugins is the industry's answer.

One package, every client

Agent Plugins is an open, vendor-neutral standard for packaging reusable components into portable plugins. Its v1.0.0 specification defines a shared format for Agent Skills and MCP servers that compatible clients can discover and load consistently. The Technical Steering Committee includes core maintainers from Amazon, Cursor, Microsoft, OpenAI, and Vercel.

The launch clients are:

• ChatGPT and Codex (OpenAI)
• GitHub Copilot (Microsoft)
• VS Code (Microsoft)
• Cursor
• Kiro (AWS)

The standard defines the minimum shared structure so the portable parts stay portable, while each client retains full control over installation, permissions, and UX.

What's inside a plugin

An Agent Plugin is a directory. The structure is intentionally minimal:

my-plugin/
├── plugin.json        # identity + spec version
├── skills/
│   └── summarize/
│       └── SKILL.md   # instructions for the agent
├── mcp.json           # MCP server config
└── com.example.client/ # client-specific extensions

Two portable component types live inside:

• Agent Skills -- instruction packages that teach an agent how to perform a specific task. At minimum, a skill is a folder containing a SKILL.md file with a name, description, and instructions. Skills can also bundle scripts, reference materials, and templates.
• MCP servers -- configuration for connecting to external tools and services. MCP (Model Context Protocol) is an open specification for linking AI clients to external tools and data. A plugin includes an MCP server when it needs live information, actions, or service integrations. If a plugin only provides instructions and resources, MCP is optional.

Client-specific behavior goes in reverse-domain extension namespaces such as com.cursor.client/, leaving the portable core untouched.

Agent Skills: the quieter half

Agent Skills originated as an open standard created by Anthropic and released in December 2025. Governance has since moved to a cross-platform body at agentskills.io. Agent Plugins brings that format into a unified, distributable package alongside MCP.

MCP handles tools and APIs well. Agent Skills covers knowledge, workflows, and prompt-based capabilities that don't require a running server. A skill is a reusable playbook: a code review checklist, a deployment runbook, a data analysis pipeline. It packages procedural knowledge and team-specific context into version-controlled folders that agents load on demand.

A minimal skill looks like this:

---
name: pr-review
description: Review a pull request for correctness and style.
---
Check for: naming conventions, test coverage, and breaking API changes.
Flag any issues as inline comments with suggested fixes.

What each client supports

VS Code, Cursor, GitHub Copilot, and Kiro support both Agent Skills and MCP with stdio, Streamable HTTP, and legacy SSE transports. ChatGPT and Codex support Agent Skills and MCP with stdio and Streamable HTTP, but not legacy SSE. The differences are minor at v1.0 -- all five clients load the same plugin.json and skills/ directory.

Where this fits in the broader shift

In 2024, building an agentic app meant per-vendor integration: a custom plugin format for ChatGPT, a custom tools API for Claude, a custom extensions API for Gemini. By 2026, MCP has become the common integration layer across all of them. Agent Plugins extends that convergence by standardizing how you package and distribute those integrations.

Skills also reflect a broader architectural shift. Rather than monolithic agents with hard-coded capabilities, the industry is moving toward composable agents that assemble abilities from portable skill packages, starting minimal and acquiring expertise as needed.

How to build your first plugin

The spec is available at agent-plugins.org, openly licensed and developed in public on GitHub. Building and distributing plugins is free.

1. Create a directory with a plugin.json manifest pointing to the v1.0.0 schema.
2. Add a skills/ directory with one subdirectory per skill, each containing a SKILL.md.
3. Optionally add mcp.json to wire up an MCP server via stdio, Streamable HTTP, or SSE.
4. Drop the plugin directory into a compatible client or follow that client's distribution instructions.

Administrators can manage plugins through policy settings and private marketplaces, distributing, permitting, or blocking specific plugins across teams. That makes Agent Plugins viable for enterprise teams standardizing internal agent tooling across multiple IDEs and products, not just individual developers sharing utilities.