Google Joins Agent Plugins 1.0 as a Core Maintainer

Google announced on Aug. 6, 2026, that it is joining the Technical Steering Committee for Agent Plugins 1.0.0, an open, vendor-neutral specification for packaging AI agent extensions into portable containers. Google becomes the sixth core maintainer of the project, alongside Amazon, Cursor, Microsoft, OpenAI and Vercel, and is represented on the committee by Kevin Hou, senior staff manager at Google DeepMind.

The announcement also marks the first day two Google products formally support the format: Agents CLI, which packages Google’s expert skills for agent building and operations, and Data Agent Kit, which connects AI coding agents to Google Cloud data services including BigQuery, Spanner and Cloud SQL.

What the Spec Actually Does

Agent Plugins 1.0.0 solves a mundane but genuinely painful problem: until now, every AI client that wanted to support agent extensions invented its own packaging format. A developer who wrote a useful tool — say, a script that queries a database and formats the results into a weekly report — could ship it to one client without much trouble. Trying to ship it to a second client meant forking the package and maintaining two near-identical copies that diverged over time, because each client expected different directory layouts, different manifest metadata shapes, and different ways of declaring network transports.

The fix is deliberately minimal. A plugin is a directory. Inside it, a plugin.json manifest carries exactly two substantive fields: a name and a version. Agent Skills live in a skills/ subdirectory. MCP server declarations go in mcp.json, with every entry required to carry an explicit transport type — stdio, Streamable HTTP, or legacy HTTP+SSE — so clients never have to guess. Client-specific extensions get their own reverse-domain namespace (for example, com.example.client/) that other clients simply ignore.

Google’s announcement is explicit that version 1.0.0 is a packaging format and nothing more. It defines no install mechanism, no distribution protocol, no permission model, no sandboxing, and no signature or provenance verification. Those are all listed as future work. For developers connecting plugins to production systems, that gap is real and worth tracking before deploying widely.

The Competitive Landscape

Agent Plugins sits on top of two standards that have already achieved near-universal adoption. The Model Context Protocol, released by Anthropic in late 2024, became the default way agents connect to external software. The Agent Skills specification, also from Anthropic and published in December 2025, spread to 32 tools from competing companies — including Gemini CLI, JetBrains’ Junie, AWS’s Kiro and Block’s Goose — within about three months of launch. Agent Plugins is now the container that bundles both together into one portable unit.

The notable absence on the maintainer list is Anthropic itself. Despite originating both MCP and Agent Skills, Anthropic is not among the six core maintainers of Agent Plugins 1.0.0, and Claude Code — which has its own plugin system with skills, agents, MCP servers and hooks — is not listed as a compatible client at launch. Anthropic launched its own Claude Code plugin directory in early 2026, which listed 256 entries as of mid-July. Whether that parallel ecosystem eventually converges with the Agent Plugins standard is an open question.

Why Students and Early-Career Developers Should Pay Attention

For anyone building agentic tools — a computer science student at a hackathon, a recent grad automating workflows at a first job — the practical change is immediate. Before this spec, choosing to support Cursor meant not easily supporting VS Code Copilot, and vice versa. The maintenance cost of covering both was enough to make most student projects pick one and stop there.

Agent Plugins 1.0.0 changes that calculus. A well-scoped tool — a plugin that queries a university research database, pulls together a literature summary, or automates a reporting pipeline — can now reach users on any compatible client from a single directory. The barrier to that distribution is a two-field JSON file.

The open governance structure also matters for early-career developers specifically. The project’s maintainers, contribution process and technical decisions are public, meaning no single company’s product roadmap controls the format’s direction. A student who contributes to the spec, publishes a well-built plugin, and lists that work publicly has a portfolio artifact that is legible across the exact tools — Cursor, VS Code, Gemini CLI — that employers and internship evaluators are actively using.

The security caveats in v1.0.0 are worth keeping in mind: no sandboxing and no permission model means plugins that touch sensitive or production systems should be treated carefully until future versions address those layers. But for the common student use case — building a useful tool and getting it in front of people — the spec removes what was the single most frustrating friction point in agentic development.

Source: Google AI