agentpluginsdirectory

Verified

How to Install Agent Plugins

You install an Agent Plugin through your client. The Agent Plugins 1.0.0 specification defines the package format and leaves installation out of scope, so each client ships its own route: a marketplace, a Git URL, or a local folder. ChatGPT, Codex, Cursor, GitHub Copilot, VS Code, and Kiro all load the same portable package.

What is an Agent Plugin?

An Agent Plugin is a directory on disk with a plugin.json manifest at its root. The specification puts it plainly: "A plugin is a directory rooted at a single filesystem location."

Two component types are portable in version 1.0.0, and both live at fixed locations:

  • Skills in skills/. Each immediate child directory containing a file named exactly SKILL.md counts as one skill. Clients are forbidden from searching deeper.
  • MCP servers in a root mcp.json, covering stdio, Streamable HTTP, and legacy HTTP+SSE transports.

Anything else a client wants to ship goes in a reverse-domain directory such as com.example.client/, which other clients ignore. Hooks, custom agents, slash commands, and LSP configuration never made it into the portable core, so a plugin carrying them gets its extras dropped everywhere except the client that owns that namespace.

The manifest requires two fields: $schema, pinned to https://agent-plugins.org/schemas/1.0.0/plugin.schema.json, and name. Ten top-level fields are permitted in total and the schema rejects everything else. If you are packaging your own, start at how to build an Agent Plugin.

Why there is no single install command

The specification authors drew the line at the package format. The client implementer guide states that the spec does not define "installation sources, registries, or marketplaces" and does not define "enablement, update, or cache user experience." The plugin author guide repeats it: "Client-managed installation, distribution, enablement, updates, and user interface are outside the portable specification."

That leaves no npm install equivalent and no official registry. What the spec does guarantee is that the package you install behaves the same once loaded. A conformant client "loads an Agent Plugin from a directory, validates plugin.json, and supports at least one component type: skills or MCP servers."

How installation works across clients

Every client route reduces to one of three patterns.

  1. Marketplace or directory. The client ships registries you browse in-product. Copilot CLI arrives with copilot-plugins and awesome-copilot already registered. ChatGPT and Codex share one directory that OpenAI reviews before publication.
  2. Git URL or repository. VS Code, Kiro, and Copilot CLI accept a repository address and clone the plugin themselves.
  3. Local folder. Useful while you are writing a plugin. Cursor reads ~/.cursor/plugins/local, VS Code takes paths through the chat.pluginLocations setting, and Kiro imports a directory through its powers panel.

One detail catches people out in every client: skills and MCP servers register when a session starts. Codex documentation says to "start a new session before using its bundled skills or tools," and the same holds elsewhere.

Which clients support Agent Plugins?

The specification maintains a compatible clients page listing five entries. All five load Agent Skills and MCP servers. They differ on MCP transport:

Client Skills MCP transports
VS Code Yes stdio, Streamable HTTP, legacy SSE
Cursor Yes stdio, Streamable HTTP, legacy SSE
GitHub Copilot Yes stdio, Streamable HTTP, legacy SSE
ChatGPT & Codex Yes stdio, Streamable HTTP
Kiro Yes stdio, Streamable HTTP, legacy SSE

ChatGPT and Codex are the exception on legacy HTTP+SSE. A plugin whose mcp.json declares an sse server will lose that server there while its skills keep working, because the spec isolates component failures.

Install guides by client

What the verified index shows about installable plugins

According to agentpluginsdirectory.com's verified index, 2026-08-07, 1,059 conformant manifests exist across 145 GitHub repositories. Two auto-generated collections account for 743 of them and 59 more are conformance test fixtures, which leaves 257 distinct plugins. Between them they package 1,209 skills and 83 MCP server configurations.

Composition matters when you are deciding what a plugin will actually do after installation. 93% of distinct plugins ship at least one skill, 25% ship an mcp.json, and 13 ship neither, carrying only metadata or client-specific extensions. Six manifests across four repositories still point $schema at the pre-rename open-plugins.com identifier, which strict validators reject.

Check a plugin before you install it

Installing a plugin gives its skills a voice in your agent's context and, for stdio MCP servers, runs a process on your machine. Read mcp.json before you install anything that declares one. The spec hands the client two environment variables, PLUGIN_ROOT and PLUGIN_DATA, and requires that every package-supplied path resolve inside the plugin root, so a well-behaved client will not let a plugin reach outside its directory. Trust policy and sandboxing are left to each client.

Paste any manifest into the validator to check it against the official 1.0.0 schema, or browse the full directory to see what a plugin contains before it touches your machine.

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.

Verified on 2026-08-07.