Agent plugins just became an open standard. On August 6, the multi-vendor group behind Agent Plugins 1.0.0 published the spec, and it’s the closest thing the agent world has to a universal package format. If you build agent extensions, this is the release that decides how you package them for the next year.
The launch is live on Vercel’s blog, and Google’s developers blog confirms the date. Two founding details matter here. The spec was refined by representatives from Amazon Web Services, Anysphere, GitHub, Microsoft, OpenAI, and Vercel, and the initial Technical Steering Committee holds core maintainers from AWS, Cursor, Microsoft, OpenAI, and Vercel. Google is joining as a core maintainer too, with support already shipping in its developer tools. The news landed the same day OpenAI made ChatGPT’s free tier unlimited, and the day’s AI roundup put both at the top of the list.
What agent plugins actually are
An agent plugin is a directory. That’s the whole trick, and the restraint is the point.
A plugin.json manifest identifies the plugin and the spec version, skills live in a skills folder using the existing Agent Skills format, and MCP servers get declared in mcp.json with an explicit transport type. Clients that support the format scan those fixed locations and load what they find. Components fail independently, so one broken MCP server doesn’t take the skills down with it.
The keyword here is portable. Agent Skills and MCP were already portable on their own. What never moved between clients was the box around them, because every client invented its own directory layout, manifest shape, and MCP configuration style. Agent plugins standardize that box and leave installation, distribution, policy, and the user experience to each client. The spec deliberately defines no install mechanism, no permission model, and no trust verification. Those are named future considerations, not quiet omissions.
Why the box matters more than it sounds
This is the part that’s easy to miss. The components were never the problem. The manifest was.
Every extension author has shipped the same skill or MCP server to two clients and watched the packaging diverge. One client wants different metadata, another guesses the transport from the config shape, and suddenly you’re maintaining two copies of components that were never different. It’s the pre-package-manager era of JavaScript, before package.json gave npm, yarn, and pnpm one format to share, and the analogy holds because the fix is the same: one format, many tools.
Agent plugins fix that specific failure. Package once, and the plugin carries to every compatible client. Vercel initiated the proposal, but the governance is public, the contribution process is open, and no single company’s roadmap sets the direction. That’s what makes this different from yet another proprietary plugin format, and it’s why the big names actually showed up.
The momentum is the real story
Standards live and die on adoption, and this one launched with real weight behind it.
At launch, agent plugins work across ChatGPT and Codex, Cursor, GitHub Copilot, Kiro, and VS Code. AWS made its Agent Toolkit compatible and shipped support in Kiro at launch, with more than 30 curated skills bundled across its plugins. Google’s Agents CLI and Data Agent Kit support the format today, and the company says more products are coming. When five competitors co-maintain a packaging standard and ship it the same week, that’s the ecosystem talking, not a press release.
This also lands at the right moment. Agents are finally doing real work, and real work needs real tooling. I’ve been watching ChatGPT Work run multi-hour agent sessions since July, and the missing piece was always the plumbing around the model. A portable format for skills and tools is exactly the kind of boring infrastructure that makes ambitious agents practical instead of demos.
What’s still missing
The honest version of this story has a second half.
Agent plugins solve packaging and nothing else. There’s still no distribution protocol, no permission model, and no provenance verification in the spec. Enterprise buyers will want approval workflows and security controls before plugins install themselves, and none of that ships in the format. The format also covers only two component types, skills and MCP servers. Hooks, sub-agents, and custom commands stay client-specific until the committee agrees they’re portable enough.
That’s the right call for a first release, and it’s still a gap. The hard parts of the plugin economy, trust, discovery, and governance, are all ahead of the format, not behind it. Nothing in the format blocks any client from shipping its own extension namespace, which is how the ecosystem stays honest while the standard matures.
Bottom line
This is the most consequential agent standard news of the week, and it barely registered against the free-tier headline.
I covered the best AI coding agents of 2026 back in July, and the fragmentation was already the loudest complaint. Every agent had its own skills format and its own tool config, and builders paid the tax in duplicated work. Agent plugins are the first credible attempt to end that, with the right companies in the room and the right scope discipline.
Don’t overhype it. It’s a package format, not a platform. But package formats are how ecosystems compound, and this one has the backing to actually stick. If you build agent extensions, read the spec and package once. The clients are already waiting.
