OpenAI Agent Plugins Connects ALL Harnesses (even Claude Code)
One plugin folder, six coding agents
OpenAI dropped agent plugins this morning, and it fixes a real pain: your custom skill for Claude Code won't run in Codex, Cursor config is useless in GitHub Copilot, and you keep rebuilding the same MCP servers for every tool. It's like owning six phones with six different chargers. Agent plugins make one shared folder work across ChatGPT, Codex, Cursor, GitHub Copilot, VS Code, and Kiro.
The standard itself is nothing fancy. A plugin is just a folder with three things: plugin.json (naming and metadata), skills files (instructions that teach the agent a task), and mcp.json (connections to external tools). Think of it as NPM for AI configs: build once, version it, install it anywhere. This cuts out the copy-paste drift where each setup slowly becomes its own messy fork.
The awkward part: Anthropic isn't at the table
OpenAI, AWS, Microsoft, Cursor, GitHub, and Vercel all agreed on this one format. But the two core pieces, agent skills and MCP, were invented by Anthropic. And Anthropic is not on the list. People on X are already asking if they'll join. Some say no, they'll push their own standard like they did with claude.md. In practice, that may not matter much: most harnesses read every .md file anyway, so even a competing standard would probably stay compatible with ChatGPT and OpenAI tooling.
One creator, Automation Labs, got the same plugin folder running in both Claude Code and Codex with a minor tweak. We haven't tried it ourselves yet, but that's a good sign. It means the "USB-C for AI agents" pitch may actually hold even if Anthropic never signs on. We'll test this across Claude Code, Codex, and Cursor soon, and we'll share what breaks. If you want to experiment before then, check out that video and drop a comment here.
What version 1.0 does and doesn't do
Agent plugins v1.0 is narrow on purpose. It standardizes packaging and discovery only. Marketplaces, permissions, runtimes, authentication, and distribution are still up to each client. That means some tools don't auto-load plugin MCP servers yet, so you may need a manual step or two. Support is uneven day one, but the foundation is there.
If you run more than one coding agent, try agent plugins now. You'll stop maintaining parallel configs, your custom skills stay consistent across tools, and you don't lose hours to tiny deviations when starting from scratch. We'll build a plugin from nothing and run it everywhere in a follow-up. For now, the practical move is: pick your most-used skill, package it once, and see which of your harnesses accept it without edits.