A single packaging format for agent extensions

Agent Plugins 1.0.0 is now publicly available as an open, vendor-neutral standard for extensions that add capabilities to AI agents. The specification defines a common directory format — a plugin.json manifest plus fixed locations for components — so that compatible clients can discover and load the same package without client-specific repackaging.

The standard targets two component types that already enjoy broad adoption: Agent Skills, which bundle reusable instructions and resources, and MCP servers, which connect agents to external tools and services. Both have their own specifications; Agent Plugins does not redefine them but supplies a shared contract for how clients locate these components inside a distributable plugin.

What the format specifies

A valid Agent Plugin is a directory containing a minimal JSON manifest at its root. The manifest records the specification version and the plugin name:

{
  "$schema": "https://agent-plugins.org/schemas/1.0.0/plugin.schema.json",
  "name": "my-plugin"
}

Those two fields are the only required manifest entries; the rest of the contract is expressed through the directory layout itself:

my-plugin/
├── plugin.json
├── skills/
│   └── summarize/
│       ├── SKILL.md
│       ├── scripts/
│       └── references/
├── mcp.json
└── com.example.client/

Clients that support Skills look for them under skills/, while clients that support MCP servers read configuration from mcp.json. A client may support either component type or both. Validation happens independently per component after the manifest is checked, so a single invalid component does not disable the others. For plugin authors, this eliminates the need to adapt identical Skills or MCP servers to multiple client conventions.

Deliberately narrow scope

The specification intentionally limits its portable surface to Agent Skills and MCP servers. Other extension points — commands, hooks, and agents — remain client-defined. New component types may be added in future versions by the Technical Steering Committee, but only once semantics converge and a clear portability need emerges. The narrow boundary keeps implementation effort low and gives the ecosystem room to mature.

Clients retain flexibility through a namespaced extension mechanism for client-specific data and files. Extensions stay outside the portable contract: each client defines its own namespace, and other clients ignore it. This keeps proprietary behavior from leaking into the shared format while still allowing client-specific capabilities to evolve until there is consensus to standardize them.

Governance and availability

Vercel initiated the proposal, which was refined into 1.0.0 with representatives from Amazon Web Services (AWS), Anysphere, GitHub, Microsoft, OpenAI, and Vercel. The initial Technical Steering Committee includes Core Maintainers from AWS, Cursor, Microsoft, OpenAI, and Vercel. The project is openly licensed, with public maintainers, contribution processes, and technical decisions.

Support for Agent Plugins is available at launch in ChatGPT and Codex, Cursor, GitHub Copilot, Kiro, and VS Code. Plugin authors package components once, and the same plugin loads across all supporting clients.

The specification, JSON Schemas, and guides for both plugin authors and client implementers are published at agent-plugins.org, with governance and contribution details in the Agent Plugins specification repository on GitHub. The conformance checklist in the spec defines the minimum requirements for clients that discover and load Agent Plugins.