Google Cloud has published a detailed breakdown of Agent Plugins, an open, vendor-neutral packaging standard for AI agent components, arguing that plugins are the natural next step for developers already building Agent Skills.
The Problem Agent Plugins Solves
Agent Skills already give agents reusable instructions and resources, and MCP already connects agents to tools and services, with both formats portable on their own. The gap wasn’t the individual components, it was having a consistent way to package them together, alongside client-specific behavior, so a bundle built once could travel across different AI coding agents without being rebuilt for each one.
What’s Inside a Plugin
A portable Agent Plugin package keeps its Agent Skills inside a skills/ directory, formatted according to the Agent Skills specification, while an mcp.json file describes any stdio, Streamable HTTP, or legacy HTTP+SSE MCP servers the plugin needs. For anything that doesn’t fit the shared, portable core, a reverse-domain namespace such as com.example.client/ acts as an extension space owned entirely by one client for hooks, agents, commands, or other client-specific additions, which clients that don’t recognize it simply ignore.
Google frames the goal plainly: components that belong together should be able to travel together, without forcing every client to converge on identical features.
An Industry-Wide Standard, Not a Google Format
Agent Plugins isn’t a Google-only initiative. Vercel publicly released Agent Plugins 1.0.0 on August 6, 2026, after developing the proposal together with AWS, Anysphere, GitHub, Microsoft, and OpenAI. Its initial Technical Steering Committee includes core maintainers from Amazon, Cursor, Microsoft, OpenAI, and Vercel, with proposals and technical decisions handled publicly and open to the broader ecosystem.
Google’s Own Plugins Built on the Standard
Two Google products already support the format: the Agents CLI packages Google’s expert skills for agent building, evaluation, deployment, observability, and publishing, turning coding agents like Antigravity, Gemini CLI, Claude Code, or Cursor into specialists at agent building and operations, while Data Agent Kit provides a collection of plugins that bring Google Data Cloud capabilities directly into a developer’s preferred AI coding agent or IDE.
Beyond Google’s own products, Firebase has also published agent skills that can be installed directly through Antigravity, Gemini CLI, Claude Code, Cursor, and GitHub Copilot, each using that client’s native install command.
When You Actually Need a Plugin
Google is careful to frame this as a targeted solution rather than a mandatory upgrade path. If a developer is shipping a single MCP server to a single client, a plain mcp.json file is still the simpler answer, and a single skill doesn’t need a plugin at all — Agent Plugins earns its keep specifically when components belong together and need to travel together.
It’s also worth noting what the spec deliberately leaves out. Agent Plugins v1 is a package format and nothing more: it defines no install mechanism, no distribution protocol, no permission model, no sandboxing requirements, and no trust or provenance verification.
Where to Learn More
Developers can explore the full specification, schemas, and governance process through the Agent Plugins specification repository, where proposals and feature discussions happen in the open via GitHub Discussions
