Five Big Tech Companies Agree on One Format for AI Agent Add-Ons. Anthropic Is Not Among Them
Amazon, Cursor, Microsoft, OpenAI and Vercel launched Agent Plugins, a shared package format for AI agent extensions. It bundles Agent Skills and MCP servers into one folder.
If you have ever tried to extend an AI coding assistant, you know the annoyance: every tool wants its own folder layout, its own config file, its own install ritual. Build something useful for one, and you get to rebuild it for the next. Amazon, Cursor, Microsoft, OpenAI and Vercel have now agreed on a shared fix. It is called Agent Plugins, and it is an open standard for packaging AI agent extensions so the same bundle works everywhere.
The idea is deliberately unambitious, which is probably why it got five companies to sign. A plugin is just a directory with a manifest file named plugin.json. Version 1.0.0 of the spec supports two kinds of content. Agent Skills are reusable instructions and workflows, essentially written guidance that tells an agent how to do a specific job. MCP servers are connectors that plug an agent into tools and data, using the Model Context Protocol. Both already existed as open standards. What is new is the wrapper around them.
What the standard explicitly does not cover is just as telling. There is no marketplace, no permission model, no runtime environment in version 1.0.0. It is packaging and discoverability only, developed in the open on GitHub. That keeps the scope small enough to be adopted quickly, and it leaves the harder questions, in particular who gets to decide what a plugin is allowed to touch on your machine, for later.
The odd part of the story is who is missing. Anthropic created both the Model Context Protocol and Agent Skills, the two components the spec is built on, and released them as open standards. It also recently added its own plugin system to Cowork, its desktop tool for agentic work. Yet it is not part of the Agent Plugins group. Anthropic has not said why, and it is worth not over-reading the absence: a company can decline to join a packaging committee for perfectly ordinary reasons. Still, when the inventor of the underlying pieces sits out the packaging standard built on them, that is a detail worth watching over the next few months.
What this means for you: if you do not write code, nothing here touches you directly today, but it is a decent signal about where AI tools are heading. The plumbing is being standardised, which usually means the ecosystem is settling down and getting more useful. If you do build with AI agents, this is the practical bit: a skill or connector you write once now has a realistic path to running in several products without a rewrite. Worth knowing before you get excited: version 1.0.0 covers only how things are packaged, not how they behave or what they are permitted to access. That still varies by product, and it is the part that actually determines whether installing a stranger’s plugin is a good idea.
Sources
AMD Buys Taalas, a Startup That Bakes an Entire AI Model Into the Chip
Taalas etches a model's architecture and weights directly into silicon. That makes it extremely fast and completely inflexible. AMD wants the technology alongside its Instinct GPUs.
