Agent Plugins

Quick Take

Agent Plugins is an open, vendor-neutral standard for packaging reusable AI agent capabilities into a single, portable format. It was proposed and published by Vercel, but it’s governed jointly: the Technical Steering Committee for the 1.0.0 release includes core maintainers from Amazon Web Services, Anysphere (the company behind Cursor), GitHub, Microsoft, OpenAI, and Vercel, with Google announcing it was joining the same day. It’s more accurate to describe this as an industry standard Vercel helped start than as a Vercel-only tool.

What It Actually Standardizes

Agent Plugins doesn’t invent new capabilities. It packages two things that already existed separately, Agent Skills and MCP (Model Context Protocol) servers, into one shared, portable directory structure:

• A plugin.json manifest at the root (the only strictly required part is a schema reference and a plugin name)
• Skills organized under a skills/ folder, for clients that support Agent Skills
• MCP server configuration in mcp.json, for clients that support MCP

Compatible clients discover a plugin by checking for plugin.json, then read whichever components they support. A failed MCP server doesn’t disable the plugin’s Skills, and one invalid Skill is skipped rather than breaking the whole plugin, so a plugin degrades gracefully across clients with different levels of support.

Why Portability Matters Here

Before a shared standard, a developer who built a useful reusable capability, a Skill, a set of MCP tools, generally had to repackage it separately for every AI client they wanted it to work in: one version for Claude Code, a different structure for Codex, another for Cursor, and so on. Agent Plugins removes that duplication: package a capability once, and any supporting client can load it without the author maintaining multiple parallel versions.

At launch, supported clients include ChatGPT, Codex, Cursor, GitHub Copilot, Kiro, and VS Code, a meaningfully wide surface for a first release, and a sign that reusable agent capabilities are becoming genuinely portable between coding environments rather than locked to a single vendor’s ecosystem.

How It Connects to Skills, MCP, and Portable Context

Agent Plugins sits one layer above two ideas Ainanza already covers separately: Agent Skill, a single reusable capability, and MCP, a protocol for connecting an AI client to external tools and data. Agent Plugins is the packaging and distribution layer that makes a bundle of Skills and MCP servers move between clients as one unit, which is closely related to the broader idea of portable context: capability and configuration that travels with you instead of staying locked inside one tool.

Current Limitations

• Scope is intentionally narrow in v1.0. Only Skills and MCP servers are standardized so far. Commands, hooks, and agent definitions remain client-specific for now.
• Support varies by client. Not every listed client implements every part of the spec identically; check a specific client’s documentation for exact support.
• Very new. Announced August 6, 2026, this is a first release of a young standard. Expect the spec, and the list of supporting clients, to keep evolving.

Alternatives and Related Approaches

There isn’t a direct “alternative” to an open standard in the usual sense, but related, narrower approaches include client-specific extension systems (each AI coding tool’s own plugin or extension format before this standard existed) and manually maintaining separate Skill or MCP configurations per tool, which is exactly the duplication Agent Plugins is meant to remove.

For the underlying building blocks, see the Agent Skill and Model Context Protocol glossary entries, and What Are MCP Connectors? for how MCP itself works.