Google said on Aug. 6, 2026, that it is joining Agent Plugins as a Core Maintainer and beginning support in its own products. Agent Plugins 1.0.0 is an open, vendor-neutral specification for packaging Agent Skills and MCP servers into portable plugins. A technical steering committee of Core Maintainers from Amazon, Cursor, Microsoft, OpenAI, and Vercel published the specification, and Google DeepMind senior staff engineer Kevin Hou will represent Google in the group.

The specification addresses a packaging problem for agent components that already work separately. Agent Skills provide reusable instructions and resources, while MCP connects agents to tools and services. Google said the missing piece was a common manifest and directory structure, because different clients had been using different layouts, metadata fields, and MCP configuration shapes. That fragmentation could force plugin authors to fork packages for different clients even when the skill and MCP server were unchanged.

An Agent Plugin is a directory with fixed locations. A minimal plugin includes a plugin.json manifest with the schema URL "https://agent-plugins.org/schemas/1.0.0/plugin.schema.json" and a name such as "reports-plugin." Skills live under skills/, one subdirectory each, using the existing Agent Skills format. MCP servers are declared in mcp.json with an explicit type for every entry, so clients do not infer transport from a config object; the format supports stdio, Streamable HTTP, and legacy HTTP+SSE.

The v1 design limits what plugin.json can do. It cannot relocate components or declare them inline, and clients load available components without requiring every optional part to exist. If skills/ is absent, the client loads what remains. If an mcp.json server fails to start, the client skips that entry, keeps loading the rest, and reports the failure. Client-specific additions can go in a reverse-domain directory such as com.example.client/, which unrecognized clients ignore.

Google said plugins are useful when multiple components belong together and need to travel together. A single MCP server for one client can still use mcp.json alone, and a single skill does not need a plugin. Agent Plugins v1 defines only a package format; it does not define installation, distribution, permissions, sandboxing, trust, provenance verification, or user experience.

Google said two products support Agent Plugins as of Aug. 6, 2026: Agents CLI and Data Agent Kit. Agents CLI packages Google's expert skills for agent building, evaluation, deployment, observability, and publishing for AI coding agents including Antigravity, Gemini CLI, Claude Code, and Cursor. Data Agent Kit provides plugins for Google Data Cloud that let agents manage data assets, run queries, and deploy data pipelines, with MCP server connections to BigQuery, Spanner, Cloud SQL, and more.

Google said on Aug. 6, 2026, that it is joining Agent Plugins as a Core Maintainer and beginning support in its own products. Agent Plugins 1.0.0 is an open, vendor-neutral specification for packaging Agent Skills and MCP servers into portable plugins. A technical steering committee of Core Maintainers from Amazon, Cursor, Microsoft, OpenAI, and Vercel published the specification, and Google DeepMind senior staff engineer Kevin Hou will represent Google in the group.

The specification addresses a packaging problem for agent components that already work separately. Agent Skills provide reusable instructions and resources, while MCP connects agents to tools and services. Google said the missing piece was a common manifest and directory structure, because different clients had been using different layouts, metadata fields, and MCP configuration shapes. That fragmentation could force plugin authors to fork packages for different clients even when the skill and MCP server were unchanged.

An Agent Plugin is a directory with fixed locations. A minimal plugin includes a plugin.json manifest with the schema URL "https://agent-plugins.org/schemas/1.0.0/plugin.schema.json" and a name such as "reports-plugin." Skills live under skills/, one subdirectory each, using the existing Agent Skills format. MCP servers are declared in mcp.json with an explicit type for every entry, so clients do not infer transport from a config object; the format supports stdio, Streamable HTTP, and legacy HTTP+SSE.

The v1 design limits what plugin.json can do. It cannot relocate components or declare them inline, and clients load available components without requiring every optional part to exist. If skills/ is absent, the client loads what remains. If an mcp.json server fails to start, the client skips that entry, keeps loading the rest, and reports the failure. Client-specific additions can go in a reverse-domain directory such as com.example.client/, which unrecognized clients ignore.

Google said plugins are useful when multiple components belong together and need to travel together. A single MCP server for one client can still use mcp.json alone, and a single skill does not need a plugin. Agent Plugins v1 defines only a package format; it does not define installation, distribution, permissions, sandboxing, trust, provenance verification, or user experience.

Google said two products support Agent Plugins as of Aug. 6, 2026: Agents CLI and Data Agent Kit. Agents CLI packages Google's expert skills for agent building, evaluation, deployment, observability, and publishing for AI coding agents including Antigravity, Gemini CLI, Claude Code, and Cursor. Data Agent Kit provides plugins for Google Data Cloud that let agents manage data assets, run queries, and deploy data pipelines, with MCP server connections to BigQuery, Spanner, Cloud SQL, and more.