Agent Plugins: The Future of Agent Skills

Existing Agent Skills were useful for providing on-demand expertise to models but had limitations in deployment and integration. Because skills (SKILL.md) and tools (MCP servers) existed as separate artifacts, each client required different integration methods, forcing developers to rewrite them every time.

Agent Plugins propose an open, vendor-neutral standard that packages agent skills and their dependent MCP servers into a single portable folder to solve these issues. This creates an environment where you can "Build it once, use it everywhere."

Key features of Agent Plugins include:

• Standardized packaging: Skills and tools move together and can be loaded immediately by any compatible client.
• Independent fault tolerance: The entire skill system does not halt even if a specific MCP server fails.
• Extensibility: Supports extension namespaces for client-specific behaviors.

Implementation requires a manifest file in the root directory, specifying the servers and transport methods needed for the skill via mcp.json. Additionally, for efficient navigation, clients follow a one level deep discovery rule, searching only immediate subdirectories of the skill directory.