What Are Agent Plugins?

An Agent Plugin is a directory with a plugin.json manifest at its root that packages Agent Skills and MCP server configuration into one portable bundle. Agent Plugins 1.0.0, published at agent-plugins.org, is a vendor-neutral packaging standard, so a single plugin folder loads in ChatGPT, Codex, Cursor, GitHub Copilot, VS Code, and Kiro without per-client rewrites.

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.

What an Agent Plugin contains

A plugin is a folder on disk, not an archive format or a package registry entry. The specification fixes three locations inside it:

• plugin.json at the root. Required. It names the plugin and declares which version of the spec it targets.
• skills/. Optional. Each immediate child directory holding a file named SKILL.md counts as one skill. The spec forbids recursive search, so skills/team/writing/SKILL.md loads nothing.
• mcp.json at the root. Optional. It declares MCP servers. The spec states that MCP configuration "MUST NOT be declared inline in plugin.json or loaded from any alternative core path."

Clients that want to store their own data get top-level directories named after a reverse-domain namespace, such as com.example.client/, plus a matching key under extensions in the manifest. A client that does not implement a namespace has to ignore it without validating the contents.

Of the 257 distinct plugins in agentpluginsdirectory.com's verified index (2026-08-07), 239 ship at least one skill, 64 declare MCP servers, and 59 ship both. The skills-only shape dominates: 93% of verified plugins carry skills, 25% carry MCP servers.

The plugin.json manifest

Two fields are required. $schema must equal the literal string https://agent-plugins.org/schemas/1.0.0/plugin.schema.json, pinned as a JSON Schema const. name is 1 to 64 characters of lowercase letters, digits, hyphens, and periods, starting and ending on an alphanumeric character, with no -- or .. inside. acme.tools and my-plugin pass. My_Plugin fails.

Eight optional fields follow: version, description, author (an object of name, email, url), homepage, repository, license, keywords, and extensions. The schema sets additionalProperties: false, so those ten names are the complete top-level vocabulary. Clients report unknown top-level fields and keep loading the plugin, while other schema violations are fatal and cost you the whole plugin.

The full field-by-field walkthrough with copyable examples lives on /build, and you can paste a manifest into /validator to check it against the official schema in the browser.

Which clients load Agent Plugins

ChatGPT, Codex, Cursor, GitHub Copilot, VS Code, and Kiro read the format. A conformant client has to load a plugin from a directory path, validate plugin.json, discover components at the fixed locations, support at least one component type, and isolate failures: one broken MCP server entry must not stop the client from loading the skills that sit beside it in the same plugin.

Anthropic sits outside the group. The organizations agent-plugins.org names on its Technical Steering Committee are Amazon, Cursor, Microsoft, OpenAI, and Vercel. Anthropic's own products read the component format underneath: agentskills.io lists Claude and Claude Code among the clients that load Agent Skills, so a SKILL.md written for a plugin still works there even though the plugin wrapper does not.

Why the standard exists

Before 1.0.0, an author who wrote a skill for one agent rewrote its packaging for the next. Each vendor shipped its own folder layout, its own manifest key names, and its own MCP wiring. Agent Plugins moves that boundary: authors target one directory shape, and clients compete on runtime behavior rather than on file layout.

The design keeps portable and client-owned concerns apart. Skills and MCP servers are portable, so the spec pins their location. Anything a client wants to add goes under a reverse-domain namespace, where the spec assigns it no meaning at all.

What Agent Plugins 1.0.0 leaves out

The 1.0.0 spec ships no registry, no marketplace, no installer, and no official validator. It defines the package and the loading contract, and stops there. Other component types stay outside the v1 format until their formats converge, so portable hooks, subagents, slash commands, and LSP integrations do not exist in this version.

That gap is why a verified directory matters. GitHub code search returns fuzzy matches for the schema string, and a manifest that mentions agent-plugins.org may still fail validation. agentpluginsdirectory.com fetches every raw plugin.json and checks it against the local copy of the official 1.0.0 schema before a plugin appears in the index.

The check catches real drift. Six manifests across four repositories, including TencentCloudBase and browserbase, still declare the pre-rename https://open-plugins.com/schemas/1.0.0/ schema IDs, according to agentpluginsdirectory.com's verified index (2026-08-07). Those files fail a strict 1.0.0 validator because the canonical $schema value is const-pinned.

Who governs Agent Plugins

A Technical Steering Committee runs the project, drawing maintainers from Amazon, Cursor, Microsoft, OpenAI, and Vercel. The GOVERNANCE.md charter is blunt about the shape of that arrangement: "All governance roles are held by individuals, not organizations. No seats are reserved for specific companies." The specification and docs carry a CC-BY-4.0 license; the code is Apache 2.0. The 1.0.0 specification page still carries Working Draft status.

How big is the Agent Plugins ecosystem?

agentpluginsdirectory.com's verified index counts 257 distinct plugins across 133 repositories on 2026-08-07, resolved from 1,059 conformant manifests. The gap between those two numbers has three sources: withoneai/one-agent-plugin ships 651 auto-generated integration wrappers, github/awesome-copilot ships 92 manifest stubs whose skills live in a client extension rather than in skills/, and 59 manifests are conformance test fixtures that validate precisely because their authors wrote them to fail elsewhere. Collapsing the first two and dropping the third gives the honest count.

MIT covers 174 of the 257 verified plugins, Apache-2.0 covers 39, and 33 declare no license field at all. Per-publisher and per-license breakdowns update daily on /stats.

Agent Plugins, MCP, and Agent Skills

The three specifications stack rather than compete. MCP is the wire protocol an agent speaks to a server, Agent Skills is the format of a single instruction file, and Agent Plugins is the box both travel in. The full comparison, with a table of what each one packages and who governs it, is on agent plugins vs MCP vs skills.

To browse what publishers have shipped so far, start at the directory. To publish your own, follow /build.

Verified on 2026-08-07.