Codex plugins settle the agent extension format war

A repo of example plugins doesn't usually earn a spot on GitHub's trending page. openai/plugins did, sitting around 5,600 stars, and the interesting part isn't any single plugin in it. It's the directory tree. Every plugin carries a .codex-plugin/plugin.json manifest, a skills/ folder full of SKILL.md files, optional agents, commands, and hooks, plus a .mcp.json for bundled MCP servers. The catalog itself is a marketplace.json checked into the repo, and you install from it with codex plugin marketplace add owner/repo.

If you've built a plugin for Claude Code, you can read that layout without opening the docs. Anthropic shipped the same design last October: a .claude-plugin/plugin.json manifest, the same component folders, a marketplace JSON at the repo root, the same /plugin marketplace add grammar. OpenAI didn't hide the lineage. It ships an official plugin for Claude Code, codex-plugin-cc, packaged in Anthropic's format, so you can delegate reviews to Codex without leaving a Claude Code session. When each vendor builds extensions for the other's agent, the format argument is over.

Plugins died once at OpenAI already

Worth remembering that OpenAI has been here before. ChatGPT plugins launched in March 2023 as hosted services: an OpenAPI spec, a manifest served from your domain, a review process. They were dead within a year, folded into GPTs, which then mostly became a distribution graveyard of their own.

The 2026 version inverts every one of those decisions. A Codex plugin is files in a git repo. No hosting, no OpenAPI surface, no server to keep alive. Distribution is git clone with a catalog file on top; Codex caches installed plugins locally under ~/.codex/plugins/cache/ and tracks them by semver, git ref, or npm version range. The plugin marketplace went live in late March with more than twenty partner integrations, Slack, Figma, Notion, and Sentry among them, and OpenAI followed in June with role-specific bundles aimed at sales and data teams. The trending example repo is the third beat of that rollout: seed the pattern, let the community fork it.

The reason this version will stick where 2023's didn't is that the unit of value changed. A 2023 plugin was an API wrapper the model called badly. A 2026 plugin is mostly procedural knowledge, skills that tell the agent how your team deploys, reviews, or upgrades an Expo SDK, with tool access bundled alongside. That's the thing developers were already writing by hand in every repo's instructions file.

What's actually portable now

The convergence isn't a coincidence, and part of it is formal. Anthropic published Agent Skills as an open spec in December, stewarded at agentskills.io, and OpenAI adopted it in Codex and ChatGPT within days. By spring, most major coding agents, including Gemini CLI, Copilot, and Cursor, read the same SKILL.md files. The plugin wrapper above skills was never formally standardized, but it converged anyway because everyone copied the working design: manifest, component folders, git-repo-as-marketplace.

For anyone maintaining internal agent tooling, that split is the practical takeaway. Put the durable knowledge in skills, because that layer now travels across vendors unchanged. Keep the manifest, hooks, and marketplace wiring thin, because that's the layer you'll fork per tool: .codex-plugin/ versus .claude-plugin/, .agents/plugins/marketplace.json versus .claude-plugin/marketplace.json, hook events with different names and payloads. A platform team can check one marketplace file into the monorepo per tool and share the entire skills/ tree between them. That's an afternoon of wiring, not a porting project.

Enterprise controls are where the two vendors are actually differentiating. Codex marketplaces carry per-plugin install policy (AVAILABLE, INSTALLED_BY_DEFAULT, NOT_AVAILABLE), and workspace admins can publish plugins from ChatGPT settings. If you run a platform team, that's the feature that matters: you can make the golden-path plugin present on every engineer's machine by default instead of writing a wiki page begging people to install it.

A marketplace is a git repo, treat it like one

The trade-off nobody puts in the launch post: this distribution model is curl | bash with better ergonomics. Installing a plugin can register MCP servers and lifecycle hooks that run with your credentials, on your machine, inside an agent that edits files and executes commands. There's no sandbox between a marketplace entry and your shell. OpenAI's curated directory and Anthropic's official plugin list exist precisely because of this, but the trending-repo culture pulls the other way, toward adding marketplaces from whoever posted a cool demo. Pin to a git SHA, read the hooks file before installing, and be pickier the more credentials your MCP config carries. Prompt-injection-shaped attacks hiding in a helpful-looking skill are the obvious next abuse vector, and 1.9 million public skills are already indexed on community aggregators.

My read: this is a genuine shift, not a hype cycle rerun, because the standardization already happened at the layer developers touch. Skills are portable, plugin layouts are near-identical, and switching agents no longer means rewriting your tooling. Which is exactly why the competition has moved up the stack, to curation, enterprise policy, and default install lists. Formats stopped being the moat this spring. Distribution is the fight now, and a trending repo of examples is what opening that front looks like.