Agent Plugins 1.0 Just Shipped Without Anthropic. Your Claude Skills Library Needs a Portability Plan.

On August 6, 2026, OpenAI, Microsoft, Amazon, Cursor, Vercel, and Google shipped Agent Plugins 1.0.0, a single open packaging format meant to let one folder of instructions and tool connections run across every major AI client. Anthropic, the company that literally authored the underlying Agent Skills specification, was not in the room. If your marketing team spent the last year building a library of Claude Skills, that absence is not a footnote. It is a decision you now have to make.

What Actually Shipped on August 6

Agent Plugins bundles two things into one portable directory: Agent Skills, meaning the instructions, scripts, and reference material that tell an AI what to do, and MCP servers, meaning the live connections to your CRM, analytics warehouse, or CMS. Before this standard, packaging a skill for ChatGPT looked nothing like packaging one for Cursor or GitHub Copilot. Every vendor wanted its own directory structure and its own manifest format. Agent Plugins 1.0 promises one package that Codex, VS Code, Cursor, GitHub Copilot, and Kiro can all read without modification.

That is a real win for anyone tired of rebuilding the same prompt library five times. But notice who is missing from that compatibility list: Claude. Anthropic wrote the Agent Skills spec and published it as an open standard on December 18, 2025, then declined to join the coalition that just built a packaging layer on top of it. The standard also explicitly excludes distribution mechanisms and permission models, so a marketplace and a security layer for these packages still have to be built by each platform separately. Nobody has actually solved discovery or trust yet. They solved the folder format.

Why Marketing Teams Should Care Right Now

Most marketing orgs did not build a single skill and call it done. They built a stack: a brand-voice skill, a campaign-brief generator, a competitive-intel researcher, a reporting skill that pulls from three different dashboards. Each one took real iteration to get right, the same kind of trial and error that goes into a good prompt template or a working integration. That library is now an asset with a format problem.

If your team standardized on Claude for content generation and research, your skills are written to Anthropic's spec. They will keep working in Claude. But if half your stack lives in Cursor for the marketing engineer's internal tools, or GitHub Copilot for the developer marketing team, or a ChatGPT-based workflow for sales enablement, none of that library ported over on August 6. You are now maintaining two versions of the same institutional knowledge, one for Claude and one for everything the new coalition touches, unless you actively design against that.

end

The Fragmentation Bet You Are Making Either Way

There are two ways this plays out, and both require a decision now rather than later. One path: the coalition format becomes the de facto standard because five major vendors are pushing it into their tools simultaneously, and Anthropic eventually adds a compatibility layer to avoid losing developers. The other path: Claude's install base among knowledge workers and its head start in agentic coding keep its own spec relevant enough that both formats coexist indefinitely, the way RSS and JSON Feed both still exist. Betting the entire skills library on either outcome is a mistake you do not need to make.

The practical fix is to treat your skill content as the asset and the packaging format as a thin, disposable wrapper. Write the actual instructions, examples, and constraints in plain markdown with no platform-specific syntax baked into the core logic. Keep platform manifests, whether that is a Claude Skills SKILL.md or an Agent Plugins manifest, as a small adapter layer that points at that shared content. That way, when the winner becomes obvious, you are repackaging a folder, not rewriting a year of institutional knowledge.

end

What To Actually Do This Quarter

Do not migrate anything yet. There is no marketplace, no permission model, and no confirmed adoption signal strong enough to justify a rebuild. What you should do is stop the bleeding: freeze any new skill from being written in a way that only works on one platform's manifest format, and audit the library you already have so you know exactly how many hours of rework you are exposed to if Anthropic stays out of the coalition through year end.

The teams that lose here are not the ones who guess wrong about which standard wins. They are the ones who never separated their actual marketing knowledge from the wrapper it happened to be born in. Fix that part now, and the format war becomes background noise instead of a migration project.