Vercel published Agent Plugins 1.0.0 on August 7, 2026, a specification that packages Agent Skills and MCP servers as a single plugin folder any compatible AI coding client can read. Coverage of the release names the format’s backers as Amazon, Cursor, Microsoft, OpenAI and Vercel, a group spanning two of the largest cloud platforms, a leading model lab and two of the most-used AI coding tools. That lineup, more than the folder structure it settled on, is the actual news: format proposals in agent tooling surface constantly, and the ones that survive are decided by whether named backers ship working support, not by how clean the specification reads.
The mechanics are narrow by design. A plugin is a directory holding a plugin.json manifest (with just two required fields, a schema version and a plugin name), a skills folder for Agent Skills, and an mcp.json file for MCP server configuration. A client that only supports one component type can ignore the other, and each component gets checked on its own, so a broken skill will not knock out a working MCP server bundled in the same package. A separate, per-client folder lets any tool store extra data of its own without that leaking into the shared format, a fix for a problem plugin authors already face when the same skill or server needs different metadata for every client it ships to.
That framing matters because Agent Plugins is not a rival to MCP, Anthropic’s protocol for connecting agents to tools and services. It sits a layer above it. MCP already defines how a client talks to a tool server. Agent Plugins defines how that server, and a separate class of reusable instruction sets called Agent Skills, gets packaged, discovered and distributed as one unit across different clients. A standard for distributing a standard is usually a sign the underlying protocol has enough adoption to need packaging conventions, not a sign it has stalled.
For someone maintaining agent tooling today, the practical gain is narrower than the launch suggests. An extension author who currently rewrites the same skill’s metadata for five separate client formats gets to write it once instead. That is a real but modest efficiency, and it depends entirely on Cursor, GitHub Copilot, owned by Microsoft, OpenAI’s ChatGPT and Codex, AWS’s Kiro and VS Code keeping their plugin loaders current as the specification changes, which no announcement can guarantee.
Cross-vendor format efforts in developer tooling have a mixed record. Wide initial backing tends to erode as individual vendors extend their own client-specific namespace instead of contributing changes upstream, an option Agent Plugins explicitly allows. Its public governance and open licensing address that incentive problem on paper. Whether the five backers keep contributing after the launch coverage fades is the only signal worth tracking.
Teams standardizing internal tooling on Agent Skills or MCP servers should watch which of the five listed clients ship a working plugin loader over the next two quarters, rather than treat the 1.0.0 label as proof the format is settled.
Announced by Vercel on August 7, 2026.
