Five major AI companies are backing Agent Plugins, a new open specification that bundles an agent’s skills and tools into one portable folder, letting developers carry their setup across platforms instead of rebuilding it for each client.
Vercel posted the 1.0 draft Thursday, with the partners pledging engineers to the project. The spec builds on the Model Context Protocol stewarded by the Agentic AI Foundation and on the Agent Skills standard Anthropic created in 2025. Skills carry instructions for how an agent performs a task, while MCP handles connections to databases, APIs, and shell scripts; together they aim for write-once-run-anywhere behavior.
Each plugin is a directory holding a JSON manifest and schema, skills in a subdirectory, and MCP server configuration in an mcp.json file. The 1.0 spec deliberately stays narrow, so client-specific experiments can still live in a namespace directory. A growing list of clients has adopted the format, including VS Code, Cursor, GitHub Copilot, ChatGPT, Codex, and Kiro.
Governance will sit with an independent technical oversight committee, with engineers from Amazon, Cursor, Microsoft, OpenAI, and Vercel on the core team. Google has folded its own Agent2Agent communications protocol into the foundation and is separately drafting a specification for discovering tools on the web, part of a wider push to tame the frontier of agent development.
