Agent Plugin 1.0.0: Why a Packaging Standard Is Needed After Skills and MCP

Background and Motivation

On August 6, Vercel together with AWS, OpenAI, Microsoft, Cursor, and GitHub announced the open Agent Plugin 1.0.0 standard, which quickly attracted over a million page views. The key question the article tackles is why a new standard is required when Agent Skills and MCP already exist.

Current Landscape: MCP, Skill, and Their Limitations

MCP standardizes how an agent connects to external tools at runtime, while Agent Skills provide reusable command packages in a consistent SKILL.md format. However, neither solution bundles the skill files, tool‑connection configuration, and versioning into a single installable unit. This leads to fragmented JSON configurations (e.g., claude_desktop_config.json, ~/.cursor/mcp.json, .vscode/mcp.json) and inconsistent root fields ( servers vs. mcpServers), causing copy‑paste failures, lack of clean uninstall commands, and scattered API keys that risk accidental repository leaks.

Agent Skills unify command syntax across tools ( .claude/skills/, .github/skills/, .agents/skills/, .gemini/skills/), but sharing still relies on manually copying folders, leading to version drift.

Existing plugin formats are also fragmented: Claude Code uses .claude-plugin/plugin.json, Cursor uses .cursor-plugin/, Copilot uses .github/plugin.json, and other vendors have their own layouts.

What Agent Plugin Adds

Agent Plugin 1.0.0 introduces a packaging and distribution layer without altering the runtime. A plugin is a directory with the following structure: root/ containing a visible plugin.json (minimum required fields: $schema and name) skills/ holding files that conform to the Agent Skills specification mcp.json describing tool connections

Client‑specific extensions placed under a reverse‑DNS namespace such as com.example.client/ The standard is supported by six major clients: ChatGPT, Codex, Cursor, GitHub Copilot, Kiro, and VS Code.

Comparison Table

Agent Skills (agentskills.io) – solves command‑package consistency; shortcoming: sharing via copy, no versioning, no one‑click install.

MCP – solves runtime tool‑connection standardization; shortcoming: client‑specific install configs and no uninstall.

CLI / Manual Config – custom glue scripts; shortcoming: not portable across clients.

Agent Plugin – packages skills, connections, and config into a discoverable, installable, versioned, cross‑client unit; shortcoming: it sits on top of Skills and MCP and does not replace them.

Evolution of MCP (2026‑07‑28 Third‑Gen Spec)

The third‑generation MCP spec makes the protocol stateless by removing the initialize handshake and Mcp-Session-Id. Each request now carries its own protocol version, client info, and capabilities, and a new server‑discovery mechanism replaces the handshake, enabling true round‑robin load balancing without sticky sessions.

Stateful interactions are handled via an explicit handle token returned by the tool. Client capabilities such as Sampling and Roots are deprecated, leaving only Elicitation. Extensions become formal mechanisms: MCP Apps (web UI), Tasks (invoke‑then‑retrieve), and EMA (enterprise‑managed auth), the latter already adopted by Anthropic, Microsoft, and Okta.

For Agent Plugin, a lighter MCP means the packaged connection configuration imposes less overhead.

Ecosystem Outlook

The standard deliberately defines only the package format, leaving distribution, installation, and monetization to the ecosystem. In China, the largest Skill Hub is the former SkillHub with roughly 10 000 skills; WorkBuddy supports SkillPay. Internationally, GitHub hosts skills and Vercel runs skill.sh.

Anthropic was not part of the initial six but led the third‑gen MCP spec. Claude Code uses a hidden .claude-plugin/plugin.json, whereas Agent Plugin uses a visible plugin.json, highlighting a compatibility gap.

Engineering Guidance

Low‑cost immediate action: Re‑arrange existing skills and tool‑connection files into the plugin.json + skills/ + mcp.json layout. This carries near‑zero risk and secures the format advantage.

Wait‑and‑see moves: Monitor whether the market builds higher‑level infrastructure, whether future Agent Plugin revisions add new component types, and whether Anthropic eventually joins the plugin ecosystem.

Decision checklist: Determine if the component needs cross‑client reuse, whether it depends on client‑specific hooks or sub‑agents, and which authentication model (OAuth vs. API key) is used. OAuth‑based connections pair best with Agent Plugin’s clean packaging, leaving credential handling to MCP.

Conclusion

The Agent Plugin spec itself is small; its success hinges on ecosystem adoption. Agents are becoming a primary entry point comparable to an operating system, with Skills as the most shareable application layer, yet still limited. The author likens the relationship to iPhone (runtime), App (Agent Plugin), and App Store (distribution). The author has registered agent-plugins.cn alongside the official agent-plugins.org and notes that a personal end‑to‑end content‑generation tool (13 skills + local Node CLI) could be trivially repackaged as an Agent Plugin. The next article will examine domestic Skill Hub ecosystems.