Why OpenAI Deprecated the openai/skills Repository
OpenAI has deprecated openai/skills, its Codex Skills Catalog repository. Despite 25.7k stars, it now serves only as a reference; active contributors should move to openai/plugins.
Why the Codex skills catalog disappeared
Over the second half of 2025, OpenAI's open-source repository for Codex, openai/skills, was officially marked as deprecated — news that spread through the developer community. At its peak the repo acted as the Agent Skills catalog that agents in the Codex CLI could automatically fetch and use, and it had become quite influential, earning 25.7k stars and 1.7k forks on GitHub.
What the repository was
The core concept was Skills = a folder-scoped collection of commands, scripts, and resources. Inside Codex, users could install whichever skill they wanted with a single $skill-installer command, and skills were organized into .system, .curated, and .experimental directories that distinguished built-in, curated, and experimental offerings.
Some of the most-used skills were GitHub workflow and planning/execution helpers such as gh-address-comments, create-plan, and gh-fix-ci. The project gained attention because teams and individuals could write the workflows they used frequently in SKILL.md form once, and Codex would call them automatically.
What deprecated means here
The core of the deprecation notice at the top of the README is the directive to "use the openai/plugins repository from now on." The existing skills repository is no longer the base distribution point for new skills; it remains only as a reference catalog.
The last commit landed on June 24, 2026 (about three months ago), and the repository description now reads "Plugins directory for skills and apps." In other words, instead of sending PRs to the old skills repository as before, skill creators should move their targets over to the openai/plugins repository and ship skills as skill-only plugins — that plugin format is now the officially supported path.
Why the structure changed
From OpenAI's perspective, integrating Skills into the plugin ecosystem looks far more scalable than managing them as a standalone repository. A plugin is more than a plain prompt command — it can cover permission scopes, configuration UI, and distribution packaging, which suits team-level distribution and enterprise adoption.
In practice, Codex's skills directory is still organized into ~/.codex/skills/.system, .curated, and .experimental. That structure was kept, and only the distribution repository moved over to plugins. From an existing user's standpoint, already-installed skills don't stop working — but to find or propose new skills going forward, you need to check the new repository.
The project at a glance
| Scope | Details |
|---|---|
| Repository | openai/skills |
| Status | deprecated (reference only) |
| New repository | openai/plugins |
| Notable skills | gh-address-comments, create-plan, gh-fix-ci |
| Last commit | 2026-06-24 |
So where does that leave you?
If you want to build a custom skill for Codex today, the recommended path is to keep using the existing SKILL.md format while distributing through the plugins repository, or simply drop the skill directly into your local ~/.codex/skills/. The openai/skills repository carries real historical weight and still works well as reference material for learning how skills are structured and authored — but it is worth remembering that for new projects it is a deprecated repository.
Reference links
Related posts
Open Source & Dev ToolsECC: one open harness for multiple AI agent environments
affaan-m/ECC unifies Claude Code, Codex, OpenCode, and others under one layer. Since 2.x, session adapters, MCP inventory, and worktree management matured into a multi-agent OS.
Anthropic's Official Agent Skills: How to Use Them
The `anthropics/skills` repository run by Anthropic is the official place to load Claude's Skills system. It covers the SKILL.md format and installation commands.
스타 24만 오픈소스 AI 에이전트 '헤르메스'
Nous Research가 만든 오픈소스 AI 에이전트 '헤르메스(Hermes)'를 분석했어요. CLI와 메신저 두 갈래로 쓰는 방식, 도구·스킬·메모리·MCP로 이어지는 확장 구조, 그리고 설치부터 첫 대화까지 2분이면 충분한 시작법까지 정리했어요.
Curated, fact-checked, and edited by a single operator before publishing.