Google joins Agent Plugins: package skills + MCP once, run everywhere

Write-once packaging ends per-client reformatting — and Google just joined the maintainers already behind OpenAI, AWS, Cursor, Microsoft, and Vercel.

Nowline AUG 11 2:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Google signs on as a Core Maintainer

    Google's Agents CLI and Data Agent Kit now load Agent Plugins, and Google joins the steering committee alongside AWS, Cursor, Microsoft, GitHub, OpenAI, and Vercel. A standard that shipped without Google now reaches its agent stack too.

  • One package, every major agent client

    The same plugin loads in ChatGPT and Codex, Cursor, GitHub Copilot, VS Code, and Kiro — no rearranging files or maintaining a separate build per tool. Package your capability once and it carries across every supporting client.

  • Skills and MCP servers in one folder

    A plugin is just a directory: a plugin.json manifest, an optional skills/ folder of Agent Skills, and an optional mcp.json declaring MCP servers. It wraps two standards you already use into one predictable, portable package.

  • Build one this weekend

    Three files ship a plugin: the manifest, a skills/<name>/SKILL.md, and an mcp.json pointing at a stdio or Streamable HTTP server. Clients expand PLUGIN_ROOT and PLUGIN_DATA, so you can bundle a local MCP server with its skill and hand it to any supported editor.

  • Read the fine print

    v1.0 is an "interoperability floor" — it standardizes packaging only, leaving install, distribution, permissions, and trust to each client. One gotcha: legacy SSE MCP transport isn't supported in ChatGPT/Codex, so prefer stdio or Streamable HTTP.