Plain-English guide to version 1.0

What AreAgent Plugins?

What are Agent Plugins? They are portable folders that package reusable capabilities for AI agents in one predictable format.

Think of it as a clearly labeled carry case: the manifest says what the package is, while Skills and MCP servers can provide the instructions and tools inside.

portable package

my-plugin/

One directory. Predictable structure.

plugin.jsonrequired
skills/optional
└─summarize/SKILL.md
mcp.jsonoptional

package

One format

clients

Many runtimes

01The short answer

A shared package format, not another app store.

AI tools often invent their own plugin layouts. Authors then have to rearrange the same Skill or MCP setup for every client. Agent Plugins 1.0 creates a small, common format that compatible clients can recognize and load consistently.

The key idea: package portable parts once, while keeping client-specific behavior in clearly separated extensions.
Before

Repackage for every client

Client A / pluginClient B / packageClient C / add-on
With the standard

One portable core

agent-plugin/compatible clients
02Inside the package

Three parts. Only one is required.

The official structure is deliberately small and easy to inspect.

Required

plugin.json

The label

Identifies the plugin and the 1.0 schema it targets. The minimum manifest needs only a canonical $schema and a name.

Optional

skills/

The know-how

Contains Agent Skills built around SKILL.md. Skills teach the agent repeatable ways to complete a task.

Optional

mcp.json

The tool connections

Describes MCP servers and transports so a compatible client knows how to connect to external tools.

Does a plugin need Skills or MCP? No. A plugin may include Skills, MCP servers, both, or technically just plugin.json. A manifest-only package is valid, but it usually has little practical value until a portable component or client extension is added.

03What each part does

Instructions and tools solve different problems.

Skills help an agent follow a repeatable workflow. MCP servers let it call external tools or data. The plugin is the portable package that can carry either one.

See the full comparison

Skill

Teaches how

Example: a release-review Skill can teach the agent what to inspect, which checks to run, and how to report risk.

MCP server

Provides tools

Example: an issue-tracker MCP server can let the agent read, create, or update issues through a defined tool interface.

Agent Plugin

Packages the parts

The predictable wrapper makes portable components easier to distribute, inspect, validate, and load.

04From folder to capability

How a compatible client uses a plugin.

01

Discover

Find the plugin directory and read plugin.json first.

02

Validate

Check the schema and supported component locations.

03

Load

Expose valid Skills and MCP servers through the client.

Compatibility is still client-dependent. The standard defines a shared floor; clients still control installation, permissions, supported transports, and user experience.
05Know the boundaries

What the standard does not promise.

A standard can make packages understandable without deciding every product choice around them.

It is not a plugin store

The official project defines a format. Directories and marketplaces are separate products.

It does not define one install command

Installation and updates remain under each client or distribution channel.

Conformance is not a security certificate

A valid layout does not prove that code, servers, or behavior are safe.

Client-specific does not automatically mean v1

A package needs the official portable manifest and structure to claim Agent Plugins 1.0 conformance.

06Common questions

What are Agent Plugins in practical terms?

These short answers cover the decisions most people need to make before they browse, build, or load a plugin.

What makes a package an Agent Plugin?
The defining feature is a root plugin.json that points to a recognized Agent Plugins schema and supplies a valid name. A package can then place Agent Skills in fixed subfolders, describe MCP servers in mcp.json, or add client-owned extensions in separate namespaces. The predictable boundary is what lets people inspect the package and lets compatible clients discover its parts without guessing where everything lives.
Why does portability matter?
Without a shared format, authors may need to rename files, move folders, or duplicate configuration for every AI client. A portable core reduces that repeated packaging work. It also gives users a familiar place to look for the manifest, instructions, and tool connections. Portability does not mean every client behaves identically; it means the reusable parts start from a common, understandable structure.
Can every AI client load every plugin?
No. A client must first support the Agent Plugins version and at least one component type. It may support Skills but not MCP, or only some MCP transports. Installation, permissions, updates, and user-facing controls also remain client decisions. Always check the client's own documentation instead of treating a valid manifest as proof of universal compatibility.
How should I evaluate a plugin?
Start with the source. Confirm who publishes it, what the Skills instruct the agent to do, which MCP servers it starts or connects to, and whether it requests credentials or broad permissions. Review the repository, license, recent changes, and verified installation documentation. Format conformance only tells you that the package follows the standard; it is not a security review, quality score, or official recommendation.
Official reference example

See the format in a real repository.

The official example includes a valid manifest and one migration Skill. It is a learning aid—not an official recommendation or a marketplace listing.

Open on GitHub

Ready to explore?

Compare repositories with sources you can inspect.

View rankings