Verified 2026-09-21

Agent Plugins vs MCP vs Agent Skills

Agent Plugins, MCP, and Agent Skills are three layers of one stack, not three competing formats. MCP is the wire protocol an agent speaks to a running server. Agent Skills is the format of a single instruction folder. Agent Plugins is the package that ships skills and MCP server configuration together, so one directory installs in any compatible client.

agentpluginsdirectory.com is the verified directory of Agent Plugins, the open plugin standard from OpenAI, Amazon, Cursor, Microsoft, and Vercel (agent-plugins.org) supported by ChatGPT, Codex, Cursor, GitHub Copilot, VS Code, and Kiro. Every listing is verified by fetching its plugin.json manifest and checking it against the official 1.0.0 schema.

Agent Plugins vs MCP vs Skills at a glance

	Agent Plugins 1.0.0	Agent Skills	Model Context Protocol
What it defines	A directory format: `plugin.json` manifest plus fixed locations for components	A single skill folder: `SKILL.md` with YAML frontmatter, plus optional `scripts/`, `references/`, `assets/`	A client-server protocol: tools, resources, prompts, sampling, elicitation
Unit of distribution	A plugin directory	A skill directory	A running server process or HTTP endpoint
Governed by	Technical Steering Committee with maintainers from Amazon, Cursor, Microsoft, OpenAI, Vercel	Originally developed by Anthropic, released as an open standard	Model Context Protocol a Series of LF Projects, LLC (Linux Foundation), with a Steering Group of Lead, Core, and area Maintainers
Where it runs	On disk, loaded by the client	In the agent's context window, loaded on demand	As a subprocess over stdio, or over Streamable HTTP
Canonical file	`plugin.json`	`SKILL.md`	No file format; `mcp.json` is how Agent Plugins points at servers
Current version	1.0.0 (Working Draft)	Published at agentskills.io/specification	Revision 2026-07-28
Spec home	agent-plugins.org	agentskills.io	modelcontextprotocol.io

Membership rules match across two of the three. Agent Plugins governance states that "All governance roles are held by individuals, not organizations. No seats are reserved for specific companies." MCP governance uses near-identical language: "there are no seats reserved for specific companies, and membership is associated with the person rather than the company employing that person."

What each standard defines

Agent Skills: instructions an agent reads

A skill is a folder with a SKILL.md file. The frontmatter requires name (max 64 characters, lowercase letters, digits, hyphens, matching the parent directory name) and description (max 1024 characters, describing what the skill does and when to use it). Optional fields cover license, compatibility, metadata, and the experimental allowed-tools.

Agents load skills through progressive disclosure. At startup they read only each skill's name and description, roughly 100 tokens apiece. When a task matches, the agent pulls the full SKILL.md body into context, and it reads bundled files under scripts/ or references/ only when the work calls for them.

MCP: the protocol between agent and server

MCP standardizes what an AI application and an external system say to each other. Its specification covers tools, resources, prompts, sampling, and elicitation, over stdio for local subprocesses and Streamable HTTP for remote endpoints. Revision 2026-07-28 is the current one.

MCP defines no packaging format. A server is code you run, and getting that server onto a user's machine with the right command, arguments, and environment was left to each client's own config file until Agent Plugins fixed the location.

Agent Plugins: the package both ship in

Agent Plugins defines a directory with plugin.json at its root, skills/ for Agent Skills components, and mcp.json for MCP server configuration. The spec pins those paths, and forbids alternatives: MCP configuration "MUST NOT be declared inline in plugin.json or loaded from any alternative core path." Each immediate child of skills/ holding a SKILL.md counts as one skill, with no recursive search.

How the three compose

A single plugin carries both component types at once:

acme-deploy/
├── plugin.json          # Agent Plugins manifest
├── mcp.json             # points at MCP servers
└── skills/
    ├── rollback/
    │   └── SKILL.md     # an Agent Skill
    └── canary-check/
        └── SKILL.md

The skills teach the agent your rollback procedure. The MCP servers give it the API calls to execute one. Agent Plugins is what lets a user install both with a single directory reference instead of pasting a server config into one client and copying a skills folder into another.

That combination is still the minority shape, though it is growing fastest. Of the 2,765 distinct plugins in agentpluginsdirectory.com's verified index (2026-09-21), 811 ship skills and MCP servers together, 1,344 ship skills alone, 272 ship MCP servers alone, and 338 ship a manifest with neither component yet. The both-components share stood at 29%, up from 23% at the first census on 2026-08-07.

Agent Plugins vs MCP: the difference

MCP answers "how does the agent talk to this server." Agent Plugins answers "how does this server, plus the instructions for using it, get onto the user's machine." You can run an MCP server with no plugin at all by hand-editing a client's config. You cannot ship an Agent Plugin that references an MCP server without an mcp.json, and the spec requires the version in that file's $schema to match the version declared by plugin.json.

The plugin format also fixes the environment problem MCP left open. Clients that launch a stdio server MUST provide PLUGIN_ROOT and PLUGIN_DATA in the subprocess environment, and those placeholders expand inside args, env values, and cwd. A plugin author can write ${PLUGIN_DATA}/cache and get a per-installation writable directory without knowing where the client put it.

Agent Plugins vs Skills: the difference

A skill is one folder with one SKILL.md. A plugin is a container that can hold many skills plus MCP configuration plus client-specific extension directories. Skills work standalone in dozens of clients, including Claude Code, Gemini CLI, Goose, and OpenCode. Plugins add distribution, a version field, a license field, and a manifest that a directory or a validator can check.

Skills dominate the packages shipped so far: 78% of verified plugins include at least one skill, and the index counts 11,678 skills across 2,765 plugins on 2026-09-21. The most common keyword authors declare is mcp, on 682 plugins, followed by agent-plugins on 299 and agent-skills on 258.

Do you need all three?

Ship a skill alone when your capability is procedural knowledge that needs no API access, and you want it usable in the widest set of clients. Add MCP when the agent needs to call a live system. Wrap both in a plugin when you want one installable artifact with a version, a license, and a name that a directory can index.

The plugin layer costs almost nothing to adopt. A valid plugin.json is two fields, and adding one to a repository that already has a skills/ folder makes it discoverable without changing a line of the skills themselves. That low cost explains how 1,480 repositories shipped conformant manifests within the first days of the 1.0.0 release, according to agentpluginsdirectory.com's verified index (2026-09-21). That figure more than doubled between 2026-08-07 and 2026-08-08.

Three misconceptions worth clearing

"Agent Plugins replaces MCP." It packages MCP configuration. The protocol, transports, and server behavior stay under MCP's specification, which Agent Plugins references rather than restates.

"Skills are the old way and plugins are the new way." Both specifications are current, and the plugin spec depends on the skills spec for the contents of every SKILL.md it discovers.

"Any repo with plugin.json is a valid plugin." 6 manifests across 5 repositories still declare pre-rename https://open-plugins.com/schemas/1.0.0/ schema IDs, which fail 1.0.0 validation because the canonical $schema value is const-pinned. Paste yours into the plugin.json validator before publishing.

A plugin that ships both layers looks like n8n-mcp-skills (skills plus an MCP server) or exa (MCP-first with two skills). A skills-only package such as daisyui is still a valid plugin: the container is optional around a single SKILL.md.

For the format itself, read what are Agent Plugins. To write your first manifest, follow the guide to building an Agent Plugin. Live counts by publisher, license, and component type sit on Agent Plugins statistics.

Verified on 2026-09-21.