Vercel, OpenAI, Microsoft, AWS, and Cursor published Agent Plugins 1.0 on August 6, 2026 — an open, vendor-neutral format that packages Agent Skills and MCP servers into one portable plugin, so an extension built once runs across ChatGPT, Cursor, GitHub Copilot, and VS Code instead of being rebuilt per product.
2026-08-08 · by Moe Ameen
On August 6, 2026, a group of normally-competing AI companies published Agent Plugins 1.0.0, described by Vercel as "an open, vendor-neutral standard for plugins that extend AI agents." The proposal was initiated by Vercel and refined collaboratively by representatives from Amazon Web Services, Anysphere (the maker of Cursor), GitHub, Microsoft, OpenAI, and Vercel. AWS, Cursor, Microsoft, OpenAI, and Vercel sit on the technical steering committee as core maintainers. The release landed the day before GPT-5's first anniversary — GPT-5 launched August 7, 2025 — a coincidence tech press framed as OpenAI looking beyond individual models toward a shared agent ecosystem.
The problem it solves is fragmentation. Over the past year two building blocks became common across agent tools: Agent Skills (reusable sets of instructions that teach an agent how to do a task) and MCP servers (the Model Context Protocol connections that give an agent live access to tools and data). But every product packaged them differently, so an extension written for one agent had to be rebuilt to work in another. Agent Plugins defines one container: a plugin.json manifest at the root, Agent Skills discovered under a skills/ directory, and MCP servers configured via mcp.json. A client can support either component or both. The pitch is "build once, run anywhere."
The standard is deliberately narrow. It specifies only how a plugin is packaged and discovered — it leaves installation, distribution, policy, and the finished user experience to each client. Six clients support the format at launch: OpenAI's ChatGPT and Codex, Cursor, GitHub Copilot, Kiro, and VS Code.
For creators, none of this is a consumer app you sign up for — it is plumbing for the developer tools that build AI agents. But it signals where the ecosystem is heading: the tools an agent can reach are converging on shared standards, which makes agents more portable and more capable orchestrators over the next year.
Read the fine print and the opportunity is obvious: Agent Plugins standardizes how an agent's *tools* are packaged, but it says nothing about the finished thing a creator actually needs — on-brand posts, scheduled and published across platforms. The spec explicitly leaves installation, user experience, and distribution to each client. That "last mile" is the whole point of [Kompozy](/). A portable plugin can teach an agent to call an API; it can't decide your brand voice, keep a persona's face consistent across a month of video, or fan one idea into [18 output formats](/glossary/output-buckets). Kompozy is the client that owns that finished output — and it already draws on Claude, OpenAI, Gemini, HeyGen, and fal.ai under the hood, so it benefits from a converging tool ecosystem without you having to wire up a single [MCP](/glossary/model-context-protocol) server yourself.
Concretely: drop in a long-form video or a single prompt and Kompozy generates [Clipped Shorts](/glossary/clipped-short), [Persona Shorts](/glossary/persona-shorts), carousels, quote graphics, a blog article, and a newsletter, then schedules them across the eight social platforms plus blog and email on [Autopilot](/glossary/autopilot). The interoperability news is a tailwind, not a threat: as agent tooling standardizes, the models and connectors underneath Kompozy get cheaper and more capable, while the part that actually earns — a distinctive voice and a wide distribution footprint — stays yours. Own the finished pipeline; let the vendors fight over the plumbing.
Agent Plugins is an open, vendor-neutral standard, released as version 1.0.0 on August 6, 2026, for packaging AI-agent extensions so one plugin works across multiple products. Vercel initiated the proposal, and it was shaped by representatives from Amazon Web Services, Anysphere (Cursor's maker), GitHub, Microsoft, OpenAI, and Vercel. AWS, Cursor, Microsoft, OpenAI, and Vercel are the core maintainers on its technical steering committee.
It defines a single container for two building blocks: Agent Skills (reusable sets of instructions) discovered under a skills/ directory, and MCP servers (Model Context Protocol tool connections) configured via an mcp.json file, all described by a plugin.json manifest at the plugin's root. Deliberately, it covers only packaging and discovery — not installation, distribution, policy, or the finished user experience, which each client handles its own way.
Six clients supported the format at launch: OpenAI's ChatGPT and Codex, Cursor, GitHub Copilot, Kiro, and VS Code. Because the format is portable, a plugin built for one of these is designed to run in the others without being rebuilt — the standard's 'build once, run anywhere' pitch.
No. The standard is plumbing for developer tools — it makes the tools an agent can call more portable, but it explicitly leaves the finished experience to each client. It does not generate on-brand posts, clip your long-form, render persona video, or schedule content across platforms. A content engine like Kompozy owns that last mile, turning any model or connector's output into finished content published across the eight social platforms plus blog and email.