Vercel drafts Agent Plugins 1.0, AAIF adopts spec for agent skills

Vercel posted the initial draft of the Agent Plugins spec on Thursday, August 6, 2026, with input from Amazon, Cursor, Microsoft and OpenAI, all of which pledged fidelity and engineers to the project. The Linux Foundation's Agentic AI Foundation (AAIF) quickly embraced the spec, though it will operate as an independent entity rather than being folded directly into AAIF's own projects.
Agent Plugins 1.0 packages an AI agent's skills and tools into a single directory structure, following the file-system-as-configuration practice Vercel already uses in Next.js. Each plugin gets a top-level directory with a small manifest that states its name, version and schema in JSON; skills live in a skills sub-directory; and the MCP server configuration is stored as an mcp.json file. Plugin authors can also use a namespace directory to store additional, non-standard extensions. Bundling the skills together with the MCP configuration is what makes the package portable across any agent client that supports the format. The spec builds on the AAIF-stewarded MCP protocol and on the open Agent Skills standard that Anthropic originally created in 2025, and it is released under the Creative Commons Attribution 4.0 International License (CC-BY-4.0).
The AAIF said it deliberately kept the scope of this 1.0 release "narrow." It is still up to individual agent clients or platforms to provide a user interface, establish security boundaries and create marketplaces where the plugins can be shared; the spec itself does not cover any of that. So far, VS Code, Cursor, GitHub Copilot, ChatGPT, Codex and Kiro all support the standard. Developers are already adding all manner of experimental hooks, rules and extensions to their skills, and for now those have to go into the unstandardized namespace directory; the best of them may be considered for future versions of the spec, though the AAIF made no promises and gave no timeline.
Going forward, Agent Plugins will be managed under an open governance model, with an independent technical oversight committee weighing public input. The core development team currently consists of engineers from Amazon, Cursor, Microsoft, OpenAI and Vercel.
Agent Plugins is not the only effort to standardize the agent landscape. Google folded its own Agent2Agent (A2A) communications protocol into AAIF and is separately building a specification called Agentic Resource Discovery for finding tools on the web. AAIF also runs a project it calls "complementary" to Agent Plugins, Skills Over MCP, which instead bundles skills directly into the MCP server itself; AAIF Vice President Angie Jones described that approach as akin to shipping "the manual with the product." How the two approaches are meant to be reconciled is left as an open question.
Key facts
- Vercel posted the initial draft of the Agent Plugins 1.0 spec on Thursday, August 6, 2026, with input and pledged engineers from Amazon, Cursor, Microsoft and OpenAI
- The Linux Foundation's Agentic AI Foundation (AAIF) quickly adopted the spec, which will run as an independent entity under an open governance model with a technical oversight committee
- The spec packages an agent's skills (built on Anthropic's 2025 Agent Skills standard) and MCP server configuration into a single directory structure, licensed under CC-BY-4.0
- VS Code, Cursor, GitHub Copilot, ChatGPT, Codex and Kiro already support the standard, though platforms still have to supply their own UI, security boundaries and marketplaces
- Google folded its Agent2Agent (A2A) protocol into AAIF and is building a separate Agentic Resource Discovery spec, while AAIF's own "Skills Over MCP" project offers a rival, "complementary" approach to Agent Plugins
Why it matters
AI agent tooling has been fragmenting: every platform wanted skills and tool integrations written in its own format. Agent Plugins 1.0 is an attempt by the biggest names building agent products to agree on one packaging format so a developer's tools and skills move between platforms without a rewrite. It slots in alongside MCP and Anthropic's Agent Skills standard as the third piece of a stack that AAIF and its backers want to make the default way agents carry their capabilities.
Who it affects
Developers who build AI agent tools and skills gain a single format to target instead of one per platform. Agent-platform vendors, including VS Code, Cursor, GitHub Copilot, ChatGPT, Codex and Kiro, which already support the standard, still have to build their own UI, security model and marketplace on top of it. The companies steering the spec directly, Amazon, Cursor, Microsoft, OpenAI and Vercel, supply the core development team, and Google is pursuing its own related but separate standardization efforts.
How to use it
A plugin is a directory: a manifest file states its name, version and schema in JSON; a skills sub-directory holds the Agent Skills content; an mcp.json file holds the MCP server configuration; and a namespace directory can hold anything non-standard. The spec is free to adopt under CC-BY-4.0. It does not by itself give a plugin a user interface, security boundary or marketplace; those remain the job of whichever client or platform runs it.
How solid is it
This is a real, published draft, not a proposal: Vercel wrote it, Amazon, Cursor, Microsoft and OpenAI contributed input and engineers, and six agent platforms already support it. The Linux Foundation's AAIF, which also stewards MCP, adopted it within days and set up an open-governance structure with a technical oversight committee around it. It is still a 1.0 release with a deliberately narrow scope, so how far actual adoption goes beyond day-one support remains to be seen.
Risks and caveats
The spec explicitly leaves out the experimental hooks, rules and extensions developers are already building; the AAIF gave no promise and no timeline for when, or whether, any of that gets standardized. AAIF simultaneously backs a second, "complementary" project, Skills Over MCP, that bundles skills into the MCP server rather than a separate plugin directory, and the article does not say how the two are meant to reconcile. Because platforms still supply their own UI, security boundaries and marketplaces, a plugin that runs on one client is not guaranteed to behave identically on another.
“Keeps the shared format predictable without requiring it to absorb every client-specific experiment happening across agent products.”
— AAIF, in a blog post