Agent Plugins 1.0: What the Open Standard Actually Means for Claude Skills
Agent Plugins 1.0.0 was published on August 6, 2026 by a Technical Steering Committee comprising AWS, Anysphere (Cursor), Microsoft, OpenAI, and Vercel. The spec lives at github.com/agentplugins/agent-plugins-spec. Codex CLI v0.147.0 shipped August 7 as the first implementation — it installs portable Agent Plugins, searches local/personal/workspace/remote plugin catalogs, and imports Cursor-managed skills. Launch clients: ChatGPT, Codex, Cursor, GitHub Copilot, Kiro, and VS Code.
Plugins do not replace skills. A plugin's `skills/` directory holds standalone skills unchanged — the manifest adds an `mcp.json` alongside it so clients discover the bundled MCP server automatically. The practical effect: skills you install today are forward-compatible with the Agent Plugins standard. The spec is currently labeled 'Working Draft' — no permission model, no sandboxing, and no signature verification are defined yet. Those are listed as future work. Ship Claude skills now; they will be valid Agent Plugins when the standard finalises.
To install a plugin via Codex CLI v0.147.0: `codex plugin install <owner>/<repo>`. For Claude Code: `claude plugin marketplace add <owner>/<repo>` still works for skills, and plugins that include an Agent Plugin manifest are handled automatically. Browse plugins at github.com/agentplugins/agent-plugins-spec to track compatible packages as they publish. The steering committee forbids any single vendor from holding a majority of seats, making this a genuinely neutral standard rather than a vendor-controlled one.
The structure of an Agent Plugin
An Agent Plugin is a directory containing three components:
plugin.json— the manifest. Declares the plugin’s name, description, version, and which skills and MCP servers are included.skills/(optional) — a folder of Agent Skills. These are the same SKILL.md files Claude Code uses today, unchanged.mcp.json(optional) — MCP server configuration. Declares which servers the plugin requires so the client can set them up automatically.
The key insight: if you have a Claude skill today and add a plugin.json to wrap it, you have an Agent Plugin. The skills inside the skills/ directory are not modified by the Plugin wrapper. A SKILL.md that works in Claude Code continues to work identically inside a plugin. The manifest is purely additive discovery metadata.
Skills vs plugins vs MCP — the decision table
| Claude Skill | Agent Plugin | MCP Server | |
|---|---|---|---|
| Format | SKILL.md markdown | Directory with plugin.json | Server with MCP protocol |
| Install target | .claude/skills/ | Plugin catalog | MCP client config |
| Requires runtime? | No — markdown only | Depends on bundled MCP | Yes |
| Works in Claude Code today? | Yes | Yes (via v0.147+ compatible clients) | Yes |
| Portable across agents? | Via plugins | Native portability goal | Via MCP standard |
| Sandboxed? | No | Not yet (future work) | Depends on server |
The practical question for a builder: do I build a skill, a plugin, or an MCP server?
Build a skill if your workflow is instruction-based — behavioral rules, task sequences, code style guides. No runtime required, installs in seconds, works today across Claude Code, Cursor, and Codex.
Build a plugin if your skill needs a bundled MCP server (API access, persistent state, tool calls beyond the agent’s built-in tools). The plugin manifest ties the skill and the server together so users install one package.
Build an MCP server if you need persistent state, external API calls, or capabilities beyond what markdown instructions can provide — and you’re comfortable with the operational complexity of running a server process.
What the launch clients tell you about adoption
The six launch clients — ChatGPT, Codex, Cursor, GitHub Copilot, Kiro, VS Code — represent the majority of agentic coding sessions happening today. This is not a fringe standard. When OpenAI, Microsoft, and Vercel coordinate on the same package format, the rest of the ecosystem follows within 6–12 months. Claude Code is the notable absence from the launch client list — watch for Anthropic’s compatibility announcement.
What the Working Draft status means
The 1.0.0 spec is labeled “Working Draft,” not a frozen final standard. Specifically absent:
- No permission model — a plugin can request any capability; there is no standardized way for the client to gate permissions
- No sandboxing — skills and servers run with the same access as the installing agent
- No signature verification — you cannot cryptographically verify a plugin’s source today
- No secrets mechanism — API keys and tokens are not standardized in the spec
These are listed explicitly as future work. For builders: treat the current spec as the stable foundation (the plugin manifest format and skills directory structure will not break), but do not build security-critical workflows around the assumption that plugins are sandboxed.
The bottom line for Claude skills
Nothing breaks. Skills you install and build today are forward-compatible. The Agent Plugins standard adds a distribution layer on top of skills without changing what skills are. The 11 months from June 2025 (skills launch) to August 2026 (plugins standard) followed the same trajectory as MCP: Anthropic ships the primitive, the ecosystem standardizes the distribution.
Browse the full Claude skills directory at aicall.ink/skills — every skill listed there is a valid Agent Plugins skills/ entry today.