Radar · 07/08/2026 · business

Agent Plugins: Five Players Sign First Shared Standard for Agent Plugins

What happened.

Amazon, Cursor, Microsoft, OpenAI, and Vercel have jointly signed Agent Plugins, an open standard that defines a single format for packaging AI agent extensions. Version 1.0.0 is built on a plugin.json manifest file and supports two components: Agent Skills (reusable instructions and workflows) and MCP servers (connectors that give agents access to tools and data).

Why it matters to you.

If you build agents, you know what it means to adapt the same extension for every platform: different directories, different manifests, different setup processes. A standard shared by the five most-used platforms means a plugin written once runs on all of them. It’s the difference between writing a tool for Cursor and rewriting it for OpenAI, or writing it just once.

The standard covers only packaging and discoverability. It doesn’t touch marketplaces, permissions, or execution environments. It’s the least political part of the ecosystem, and therefore the easiest to converge on.

If you want to try it.

The specification is public on GitHub. If you have an MCP server or a set of reusable instructions, it’s worth reading the plugin.json format and understanding what’s needed to package your work in the new standard.

In detail

Until yesterday, each agent platform had its own format. Cursor, OpenAI, Vercel, Amazon: each with its own folder structure, manifest, and way to declare what an extension can do. If you built a useful tool, you rewrote it four times.

Agent Plugins is an attempt to stop this fragmentation before it becomes irreversible. Five players, three of whom already have agent platforms in large-scale production, agree on a minimal packaging format.

The format is simple: a directory with a plugin.json file inside. The manifest describes two types of components. Agent Skills are blocks of instructions and workflows that an agent can load and follow. MCP servers are connectors that expose tools and data to the agent, and which we saw become stateless and orchestrable with MCP 2.0 early this month.

The choice to build on top of MCP is significant. The Model Context Protocol is already an open standard, born at Anthropic and adopted by a growing number of providers. Agent Skills have the same pedigree: they’re an open format that originated at Anthropic. The five signatories take two things that already exist and package them in a single discovery format.

What the standard doesn’t cover is as important as what it does. No marketplace: who distributes and who installs remains outside the spec. No permissions: the format says what’s in the plugin, not who can use it. No runtime: the execution environment remains the responsibility of whoever hosts it. It’s a pragmatic choice. The points where platforms diverge most (monetization, access control, isolation) remain competitive. The points where converging is pure advantage for everyone become standard.

Anthropic’s absence is the most interesting detail. Anthropic created both MCP and Agent Skills as open standards, and just added its own plugin system to Cowork, the desktop tool for agentic work. Not participating in the initiative can mean two things: either it considers its own ecosystem sufficient, or it doesn’t want to converge on a format controlled by a group where it has no voice. Either way, the standard launches without the company that provided its foundational pieces.

For those building agents, the signal is clear. The packaging and discovery layer is normalizing, as happened with containers through OCI or browser extensions through WebExtensions. Competition shifts higher: whoever offers the simplest marketplace and the most secure runtime wins. If you have an MCP server or a skills library, starting to think in terms of plugin.json makes sense even if you wait to formally adopt the standard.

What remains open. The spec is at 1.0.0, but real adoption depends on when the five platforms integrate the format into their installation flows. A manifest file on GitHub is a start, not an ecosystem. And without Anthropic, the coverage of the agentic ecosystem has a hole at the point where MCP and Skills were born.

Type to search across course, playbooks, skills, papers…