Meet Agent Plugins: A New Open Standard That Unifies Agent Skills and MCP Servers in a Shared Format

AIDeveloper44 Team

Standardizing the AI Agent Ecosystem

As the landscape of artificial intelligence continues to shift toward autonomous agents, the industry has faced a significant hurdle: fragmentation. Until recently, AI agent clients developed proprietary plugin formats, requiring developers to rebuild or adapt their tools for every individual platform. To address this, a coalition of technology leaders has released the Agent Plugins specification, a vendor-neutral standard for packaging reusable components that extend AI agents.

Version 1.0.0 of the specification defines a portable package format designed to work across compatible clients. By establishing a "portability floor," the standard allows authors to package their tools once and have them discovered and loaded consistently across various environments. This effort is governed by a Technical Steering Committee (TSC) that includes core maintainers from OpenAI, Microsoft, Amazon, Vercel, and Cursor, signaling broad industry alignment on the future of agent interoperability.

The Problem of Platform Fragmentation

Before the introduction of Agent Plugins, developers creating tools for AI agents—such as specialized calculators, database connectors, or web searchers—were often forced to duplicate their work. A plugin designed for one agent client might require a different manifest, a different directory structure, or different communication protocols to work on another. This lack of a shared format increased the barrier to entry for developers and limited the utility of third-party extensions.

The Agent Plugins standard addresses this by separating the portable core of a plugin from client-specific implementations. While distribution, installation, permissions, and user experience remain under the control of individual clients, the underlying structure of the plugin's capabilities is now standardized. This ensures that a plugin containing a specific "skill" or a Model Context Protocol (MCP) server can be recognized by any client adhering to the specification.

Anatomy of an Agent Plugin

An Agent Plugin is defined as a directory containing a required manifest and optional components organized in fixed, predictable locations. According to the official documentation, the structure typically includes:

• plugin.json: The central manifest that identifies the plugin, its version, and the version of the Agent Plugins specification it targets.
• skills/: This directory contains "Agent Skills," which are sets of tools and logic defined by the separate Agent Skills specification.
• mcp.json: This file describes integration with the Model Context Protocol. It supports stdio, Streamable HTTP, or legacy HTTP+SSE MCP servers, allowing agents to interface with external data and services via a standardized protocol.
• Client Extensions: To allow for innovation without breaking the portable core, the standard supports reverse-domain extension namespaces (e.g., "com.example.client/"). This allows specific clients to add unique behaviors or hooks that other clients can simply ignore.

Open Governance and Development

One of the defining characteristics of the Agent Plugins initiative is its commitment to open source principles. The project is openly licensed under CC-BY-4.0 for documentation and Apache-2.0 for code and schemas. Development takes place in public repositories on GitHub, where proposals for new features and changes are debated in public discussions.

The Technical Steering Committee ensures that the standard evolves based on the needs of the broader ecosystem rather than the interests of a single corporation. By involving competitors like Amazon and Microsoft alongside OpenAI and Vercel, the project aims to maintain a neutral ground for AI infrastructure. This collaborative approach is intended to foster a robust ecosystem where developers can trust that their investments in building agent tools will remain viable across multiple platforms.

Integration with MCP and Agent Skills

Agent Plugins does not seek to replace existing protocols but rather to act as a container for them. It specifically integrates the Model Context Protocol (MCP), which has gained traction as a way to connect AI models to external data sources. By including mcp.json in the standard directory structure, Agent Plugins provides a reliable way for clients to discover and launch MCP servers.

Similarly, by adopting the Agent Skills format, the specification provides a way to package the "cognitive" tools an agent might use—such as scripts or references—into a format that any compatible client can execute safely. This modularity allows developers to mix and match different types of capabilities within a single, portable package.

The Road Ahead for Developers and Implementers

For plugin authors, the path forward involves adopting the plugin.json manifest and organizing their existing tools into the specified directory layout. For client implementers—those building the software that runs the agents—the task is to implement the loading, validation, and safe execution of these portable packages. The project provides canonical JSON Schemas to assist with validation and a conformance checklist for implementers.

As AI agents move from experimental prototypes to production-ready tools, standardization at the transport and packaging layer is a critical step toward maturity. By providing a shared language for agent extensions, the Agent Plugins project aims to accelerate the development of a more interconnected and capable AI ecosystem.

Enjoyed this?

Get more posts like this delivered to your inbox.

Subscribe