Agent Plugins

What Agent Plugins Is

Agent Plugins is a technical specification designed to bring interoperability to the AI agent ecosystem. It provides a standardized directory structure and manifest format that allows developers to package AI capabilities—specifically Agent Skills (pre-defined tasks or functions an agent can perform) and MCP servers (which provide context and data via the Model Context Protocol)—into portable units. By adopting this standard, a plugin created for one environment functions seamlessly in any client that supports the specification.

The Problem It Solves

The AI agent landscape often suffers from fragmentation, where developers must build and maintain separate versions of the same tool for different platforms. This creates friction and limits the reach of useful agent capabilities. Agent Plugins addresses this by establishing a ‘build once, run anywhere’ model, serving as a foundational layer for Agent Infrastructure (the platform layer that provides runtime, middleware, and guardrails for deploying AI agents). This approach ensures that useful capabilities can travel across different agents, unlocking broader interoperability.

How It Works

The standard relies on a predictable directory structure that AI clients can parse. Key components include:

• plugin.json: The root manifest file defining the plugin’s identity and configuration.
• skills/: A dedicated directory for Agent Skills, typically formatted using the SKILL.md standard.
• mcp.json: A configuration file for integrating MCP servers, which allow agents to connect to external data sources or tools.
• Client Extensions: Uses reverse-domain namespaces to ensure that specific client requirements do not conflict with the core standard.

Consider the standard as a common shipping container for freight. Just as a standardized container allows goods to move efficiently between ships, trucks, and trains without being repacked, Agent Plugins allows a developer to package a tool once so it fits into any compatible agent client, regardless of the underlying architecture.

Governance and Development

The standard is maintained by a Technical Steering Committee (TSC) composed of industry leaders. The founding coalition includes core maintainers from Amazon, Cursor, Microsoft, OpenAI, and Vercel. Google serves as a Core Maintainer, and GitHub participates as a contributing organization. Development is conducted openly on GitHub, where the community contributes to the evolution of the standard through public discussions.

What It Doesn’t Cover

The current version of the specification is strictly limited to the packaging format. It does not address complex agent-level concerns such as permissions, sandboxing, secrets management, or signature verification. These critical security and behavioral features are explicitly out of scope for the initial release and are deferred to future versions of the specification.

Why It Matters

By creating a shared, open standard that is simple to adopt, Agent Plugins reduces the barrier to entry for developers and increases the utility of AI agents. It allows for a more cohesive ecosystem where tools and skills are not siloed within a single platform, ultimately fostering innovation across the entire agent landscape.

Common Questions

Does Agent Plugins handle agent security or permissions?

No. The current version is focused exclusively on the packaging format. Security features like sandboxing, secrets handling, and permissions are deferred to future versions.

Can I use Agent Plugins with the Model Context Protocol (MCP)?

Yes. The standard is designed to package MCP servers alongside Agent Skills, making it a key component of the broader agent infrastructure ecosystem.

Who decides how the standard evolves?

The standard is governed by a Technical Steering Committee (TSC) consisting of representatives from Amazon, Microsoft, OpenAI, Cursor, Vercel, and Google, with development occurring openly on GitHub.