OpenAI Deprecated Its Codex Skills Repo

If you went looking for Codex skills this month and landed on openai/skills, the first thing on the page is a notice telling you to leave. The repository is deprecated. It still has 25.7k stars, 1.7k forks and 37 contributors, and it still sits near the top of search results for the term, but the last commit landed on 14 July 2026 and it was the one that added the deprecation banner. Everything current moved to a different repo with a different shape, and the docs that tell you how to get there don't say "skills" in the URL.

That gap between what ranks and what's maintained is the whole problem. Here's what actually changed, what's still worth taking out of the old catalog, and where a skill you already wrote is supposed to go now.

Key Takeaways

• openai/skills is deprecated as of a commit on 14 July 2026; the README now points at openai/plugins instead.
• Nothing was deleted. 39 curated skills and 5 system skills are still readable in the old repo, and installed skills keep working.
• Skills didn't go away, they got demoted: a plugin with a .codex-plugin/plugin.json manifest is the new unit of distribution, and a skills/ folder is one thing that manifest can carry.
• The old repo has no root LICENSE. Each skill carries its own LICENSE.txt, so check the folder you copied.
• Codex skills and Claude skills share a file format and the Agent Skills standard; they now differ sharply in how you distribute one.

What actually happened to the Codex skills catalog

OpenAI opened openai/skills on 25 November 2025 as a catalog: no releases, no tags, just folders. It accumulated 114 commits and 77 open issues before the banner went up. The README now points at openai/plugins, created 4 March 2026, and at a Build plugins guide that includes instructions for making what OpenAI calls a skill-only plugin.

Read that last phrase carefully, because it's the actual news. Skills weren't cancelled. They got demoted from a top-level thing you install into a component you package inside something bigger. A plugin is now the unit of distribution, and a skill is one of the surfaces a plugin can carry.

The old repo is worth keeping open in a tab anyway, because nobody deleted the contents. Two directories survive: skills/.curated with 39 entries, and skills/.system with five. The .system five are the ones Codex installs on its own (imagegen, openai-docs, plugin-creator, skill-creator and skill-installer), so if you've been using Codex at all recently you already have them without asking.

The curated 39 are a decent map of what OpenAI thought people wanted last spring: seven Figma skills, four Notion ones, deploy skills for Cloudflare, Netlify, Render and Vercel, gh-fix-ci and gh-address-comments for the GitHub loop, three security skills covering best practices, an ownership map and threat modelling, plus pdf, transcribe, speech, screenshot, jupyter-notebook and a migrate-to-codex skill that is now quietly ironic.

One detail that tells you how abruptly this happened: the README still explains how to install from a .experimental folder, and links to skills/.experimental/create-plan. That folder isn't in the repo anymore. The instructions outlived the thing they described by about three months.

Codex skills and Claude skills are the same word, not the same plumbing

Both ecosystems landed on "skill" meaning a folder with a SKILL.md and some supporting files, and both point at the Agent Skills open standard as the shared spec. That much genuinely overlaps, which is why a lot of skills written for one agent run fine in the other after you move the directory.

The divergence is in distribution, and it's now wide. Anthropic's side grew directories and installers around the loose-folder format; we walked through that ecosystem in what agent skills actually are and in more depth on Anthropic's version specifically. OpenAI went the other way and wrapped the folder in a manifest.

So if you're comparing the two, don't compare the file formats; they're close enough to ignore. Compare what happens when you want to hand your skill to someone else. On the Claude side that's still mostly "clone this and copy the folder." On the Codex side it's now a packaged artifact with an identity, published once to a directory that ChatGPT and Codex both read.

That difference shows up in the search data too. Checking both terms in Ahrefs this week (US, September 2026), "claude skills" pulls about 19,000 searches a month against 3,900 for the Codex term, which reads like a gap in adoption but is closer to a gap in naming: Anthropic shipped the word to consumers, OpenAI shipped it to a manifest. The people asking about Codex skills are, on the evidence of what they're asking, further along and stuck on plumbing rather than on concepts.

One asymmetry does matter in practice. A Codex plugin can ship an .mcp.json, meaning the thing you installed expects an MCP server to be answering when the agent reaches for it. That's fine while your laptop is open and the server is running; it's less fine for anything scheduled, and it's the point where a hosted setup stops being a preference. MoClaw runs the agent and its servers on a cloud computer that stays up, so a plugin that depends on something listening doesn't quietly stop working when you shut the lid.

Where Codex skills live now: inside a plugin manifest

Every plugin sits at plugins/<name>/ and needs a .codex-plugin/plugin.json manifest. Around that manifest the folder can also hold a skills/ directory with the workflows you wrote, an .app.json referencing a registered MCP server connection, an .mcp.json for a server shipped with the plugin, plugin-level agents/ and commands/, a hooks.json, and an assets/ folder.

The marketplace lives at .agents/plugins/marketplace.json and points at the standard plugins/ directory. If you log in with an API key rather than a ChatGPT account you get a different one, .agents/plugins/api_marketplace.json, which is the kind of detail that costs an afternoon when nobody tells you.

Public plugins publish once to a universal directory shared by ChatGPT and Codex. Local and repo marketplaces stay separate, and OpenAI's docs are explicit that availability varies by surface, so a plugin working in one place isn't a promise about the other.

The richer examples in that repo are worth reading before you write your own manifest: plugins/figma covers Code Connect and design system rules, plugins/notion handles planning and meetings, and there are bundles for building iOS, macOS and web apps, plus Expo, Netlify, Remotion and Google Slides.

Migrating a skill you already wrote

The good news is that the SKILL.md you wrote in March almost certainly still works. What changed is the wrapper around it, and OpenAI shipped a skill to generate that wrapper: @plugin-creator, which is one of the five that auto-installs. Point it at your existing folder and it scaffolds the manifest and marketplace entry. You can also write the manifest by hand; the docs say both routes produce the same structure, and I'd believe them given how thin the manifest is.

Two things to check before you assume a clean move. First, licensing: the old repo has no LICENSE file at its root, because each skill carries its own LICENSE.txt inside its directory. If you forked something out of .curated and never opened that file, you don't currently know what terms you're on. Second, the install verb changed. The old flow was $skill-installer gh-address-comments, then restart Codex. The new flow goes through plugin install, and anything in your onboarding docs still telling people to run $skill-installer against a GitHub tree URL is pointing at a repo that stopped moving in July.

There's a third thing, less obvious, and it only shows up on the second machine. Reinstalling a skill set is cheap; reinstalling the environment underneath it isn't. A deploy skill assumes a CLI is authenticated, transcribe assumes ffmpeg, jupyter-notebook assumes a kernel that imports what the notebook needs. Every time you migrate machines you rebuild that layer by hand. Running the agent on a hosted cloud computer instead means the toolchain is part of the box rather than part of your weekend, and MoClaw is built around exactly that split: the skill folder is yours, the machine underneath it is already standing.

The folder problem nobody has solved

Search for Codex skills right now and the top of Google isn't documentation. It's people asking each other where things go. A github/spec-kit discussion is titled, plainly, "Confusion with codex-cli on commands vs skills." An OpenAI community thread runs eighteen replies deep on what to do when skills in a large catalog overlap and the agent picks the wrong one. Neither thread has a tidy answer.

The deprecation makes that worse before it makes it better, because now there are two shapes in the wild and no automatic migration for anything you cloned. Community indexes are in the same position: composio-community/awesome-codex-skills, at 16.3k stars, last saw a push on 26 July 2026, which is twelve days after the official repo stopped.

If you maintain a skill catalog for a team, the practical move is to treat the old repo as an archive you read and the plugins repo as the one you write to, and to pin the docs URL rather than the GitHub URL in your internal notes, because docs pages get redirected when things move while repos just sit there with a banner on top, ranking.

It's also worth deciding, once, whether overlapping skills are a catalog problem or a scoping problem. The OpenAI thread mostly treats it as the former and argues about naming conventions; the replies that actually resolved anything did it by giving each project a narrower set rather than by writing better descriptions. That's easier to enforce when a project's skills live with the project instead of in one global folder every repo on your disk can see.

And if your team is spread across machines, the migration is the moment to stop syncing skill folders by hand. An agent running on a hosted computer keeps one copy of the plugin set that everyone reaches, which is a different problem from the one dotfiles solve; MoClaw handles that side, alongside your local Codex install rather than instead of it. We covered how the task-shaped Codex workflow fits together in ChatGPT Codex agent workflows.

FAQ

Is the OpenAI Codex skills repo still official?

It's still an official OpenAI repository, and it's still public with all 44 skills readable. It's also deprecated, by its own README, as of a commit on 14 July 2026. Official and current aren't the same thing here.

Are Codex skills the same as Claude skills?

The file format is close, and both reference the Agent Skills open standard, so a SKILL.md often moves between them with no edits. Distribution is where they split: Codex now wraps skills in a plugin manifest with a marketplace entry, while the Claude side stays closer to loose folders.

Do I need a paid plan to use Codex skills?

The catalog itself is a public GitHub repo, so reading and cloning costs nothing. Running them needs Codex, and OpenAI's docs note that plugin availability varies by surface and that API-key logins get a separate marketplace file, so check the current plan terms before you build a workflow on top of one.

What happens to skills I already installed?

Nothing breaks. Installed skills keep working; the five .system ones keep shipping with Codex. What stops is new curation in that repo, so anything you pull from it from now on is a snapshot of July.

Where should I publish a new skill?

Package it as a skill-only plugin and publish through the plugins repo route. @plugin-creator generates the manifest, and the create-skill docs still cover writing the SKILL.md itself.

What to take from this

The catalog didn't fail; it got absorbed. Skills as a format won so thoroughly that OpenAI stopped treating them as a product and started treating them as a field in a manifest, which is what happens to things that become infrastructure.

For anyone maintaining Codex skills today, the work is unglamorous: read the licence file inside each folder you copied, rewrite the install line in your docs, and stop linking your team at a repo whose last real commit is three months old. As of September 2026 the old catalog still outranks the thing that replaced it, and it will keep doing that for a while, so the correction has to come from you rather than from search.