OpenAI Killed Its Skills Catalog. That's Worse.

The openai/skills repository hit 26,600 stars before OpenAI deprecated it on June 22, 2026. That sounds like a product failure. It was actually a power move.

📖 Read the full version with charts and embedded sources on AgentConn →

OpenAI did not abandon skills. It absorbed them. The open catalog became a managed Plugin Directory with 185 entries, six role-specific plugins bundling 62 app connections and 110 skills, and a curated directory inside the Codex app itself. In four months, OpenAI ran the playbook every platform operator knows: open standard attracts contributors, managed marketplace captures distribution, first-party curation controls discovery.

If you built a third-party skills directory — and a dozen startups did — you just watched the landlord open a competing store on the ground floor.

View original post on X →

The Timeline Nobody Is Talking About

The speed of this consolidation is the story. Here is how fast OpenAI moved from "we adopted your format" to "we own the storefront":

December 2025 — Anthropic releases the Agent Skills open standard at agentskills.io. A skill is a folder with a SKILL.md file. Simple, portable, no vendor lock-in. Simon Willison called it Anthropic's "brilliant" mechanism — and noted that OpenAI had already adopted it quietly in both ChatGPT and Codex CLI.

January 2026 — The openai/skills repository goes live. System skills, curated skills, experimental skills. A $skill-installer command inside Codex. Open, forkable, community-friendly. The message: "build here."

March 2026 — OpenAI launches the Plugin Directory. Skills become one component inside a larger plugin format. Each plugin bundles a manifest, optional skills, and app connections. The directory starts inside the Codex app.

June 2, 2026 — OpenAI ships six role-specific plugins for Codex: Data Analytics, Creative Production, Sales, Product Design, Public Equity Investing, and Investment Banking. 62 apps, 110 skills, no coding required. The company announces Codex has over 5 million weekly users, with non-developers growing 3x faster than developers.

June 22, 2026 — The openai/skills README gets a deprecation notice: "For current Codex skill and plugin examples, use the OpenAI Plugins repository." The open catalog is dead. Long live the managed directory.

July 9, 2026 — OpenAI migrates the app directory to the Plugin Directory. Plugins become the "primary way to discover workflow capabilities across ChatGPT and Codex."

View original post on X →

That is seven months from "we adopted your open standard" to "we own the directory and you publish through us." If you have seen this movie before, you have — it is the iOS App Store arc compressed into half a year.

ℹ️ The Sherlock moment: In platform economics, "sherlocking" is when the platform owner ships a first-party version of a popular third-party feature. Apple does it to indie apps regularly. OpenAI did not sherlock individual skills — it sherlocked the entire distribution layer.

What the Community Saw Coming (and What It Missed)

When skills first landed in Codex, the Hacker News thread "Skills Officially Comes to Codex" hit 304 points and 130 comments. The excitement was genuine.

View on Hacker News →

User cube2222 praised the format: "Easy to author (at its most basic, just a markdown file), context efficient by default...can piggyback on top of existing tooling." User JamesSwift went further with the Matrix analogy: "Skills are the matrix scene where Neo learns kung fu...a database of specialized knowledge that an agent can instantly tap into on demand."

But btown spotted the problem nobody wanted to talk about: "There's no RAG on the content of Skill code...the names and descriptions in every skill's front-matter are included verbatim in your prompt." Translation: every skill you install costs you context tokens whether the agent uses it or not. That is a design choice that rewards a curated catalog over a sprawling marketplace — and guess who controls the curation now.

A later Ask HN thread — "Are most agent skills just fancy system prompts with a name?" — was blunter. One user analyzed 881 ClawHub skills and found that "about 46% scored poorly on functional depth, meaning they tell Claude what the skill is but never what to actually do." When nearly half the skills in a marketplace are vaporware, the platform owner has every justification to replace the open bazaar with a curated store.

⚠️ Contrarian Corner: The deprecation might actually prove the open standard won. OpenAI did not invent a competing format — they adopted agentskills.io and then absorbed distribution around it. The format itself (SKILL.md) still works everywhere: Claude Code, Cursor, Gemini CLI, GitHub Copilot, Amp, OpenCode, and 40+ other tools. If the format is genuinely cross-platform, third-party directories that index across vendors — not within one vendor — might actually get stronger as each platform builds its own walled garden. The middleman becomes the cross-vendor discovery layer. That is a real business. But it requires intentional repositioning, not denial.

The Three-Layer Power Play

What OpenAI executed is not unique. It is a pattern. And Anthropic is running a version of the same play.

Layer 1: Adopt the open standard. Both OpenAI and Anthropic use the Agent Skills format. The agentskills.io specification is cross-platform by design — Anthropic originated it in October 2025, and as of mid-2026 it is supported by over 40 tools. Adopting an open standard is low-cost credibility: developers feel safe because their skills are portable.

Layer 2: Ship a first-party catalog. OpenAI launched openai/skills. Anthropic has anthropics/skills with first-party skills like frontend-design (277,000+ installs), skill-creator, mcp-builder, and webapp-testing. These are not just reference implementations — they are high-quality defaults that set the bar for the ecosystem.

Layer 3: Absorb distribution into the product. This is where it gets dangerous for third parties. OpenAI's Plugin Directory lives inside the Codex app. When a user needs a skill, they browse the directory — not GitHub, not mattpocock/skills, not Composio's awesome-list. Anthropic's version is subtler: Claude Code ships with /skills built in, a $skill-installer equivalent, and the official skills are bundled with the product. Discovery happens inside the tool.

The pattern is not new. It is how every successful platform works:

Platform	Open Phase	Capture Phase	Current State
iOS	"Anyone can build apps"	App Store curation + 30% cut	Platform owner ships competing apps
Chrome	Extensions API, open Web Store	Manifest V3, restricted APIs	Google ships features that replace popular extensions
AWS	Open-source friendly	Managed versions (ElastiCache, DocumentDB)	Third-party OSS vendors forced to re-license
OpenAI Codex	openai/skills (open repo)	Plugin Directory (managed, in-app)	Role plugins, curated curation

View original post on X →

Who Is Exposed

The skills directory race we covered in April identified a half-dozen competitors. Here is where they stand now:

mattpocock/skills (28,500 stars) — Matt Pocock's personal skills dump. High-quality configurations from a respected TypeScript engineer, but tightly coupled to Claude Code workflows. Not a directory — a personal brand play. The least exposed, because it was never trying to be a platform.

Composio — Positioned as the skills marketplace connecting agents to 1,000+ SaaS apps. The most exposed. Composio's value proposition is exactly what OpenAI's Plugin Directory does: bundle app connections with skills in a discoverable catalog. When the platform owner ships the same thing, the third-party aggregator loses the convenience argument.

VoltAgent/awesome-agent-skills — A curated list of 1,000+ community skills. Open-source, cross-platform index. Moderately exposed — if it stays vendor-neutral and indexes across platforms, it survives as a discovery layer. If it drifts toward any single vendor's ecosystem, it dies.

Agensi — The commercial marketplace where developers buy and sell skills. Most interesting positioning: monetization for skill authors, which neither OpenAI nor Anthropic currently offers. But monetization only works if you control discovery, and that is exactly what the platform owners are absorbing.

agentskillexchange — An open catalog aiming for cross-platform coverage. Similar dynamics to VoltAgent.

What This Means for You

If you are building, publishing, or relying on agent skills, here is the operational takeaway:

1. Publish to the first-party directory first. The Plugin Directory for Codex, Claude Skills for Claude Code — that is where 5 million+ users discover skills. Third-party directories are supplementary reach, not primary distribution. This is the lesson every iOS developer learned in 2009.

2. Cross-platform is your insurance policy. The Agent Skills format is genuinely portable (agentskills.io, 40+ tools). A skill written as a SKILL.md folder works in Codex, Claude Code, Cursor, Gemini CLI, and more. Do not vendor-lock your skills to one platform's proprietary extensions. The format's portability is the one thing the platform owner cannot take from you.

3. If you run a skills directory, index across vendors or die. A Codex-only directory is a feature OpenAI can ship. A Claude-only directory is a feature Anthropic can ship. A cross-vendor index that normalizes skills across all platforms and gives developers a single search surface — that is a product the platform owners do not want to build, because it benefits their competitors equally. That is the surviving niche.

4. Watch for the security argument. The OWASP AI skill security blueprint and Unit 42's finding that 17% of early ClawHub skills carried malicious payloads give platform owners a legitimate reason to gate distribution. "We curate for security" is the argument that turns an open ecosystem into a walled garden — and it is an argument that is hard to argue against when the threat is real. The supply chain attack surface we covered in June is now a platform governance lever.

ℹ️ The deeper pattern: Skills are the new package managers. NPM started as an open registry; it is now owned by GitHub (Microsoft). PyPI started as community infrastructure; it now depends on corporate sponsorship. Docker Hub started open; rate limits and paid tiers followed. Every distribution layer in developer tooling follows the same arc: open, then essential, then captured. Agent skills directories are at the "essential" stage right now. The capture phase has begun.

The Anthropic Parallel

Anthropic is not innocent here. While OpenAI's move is louder, Anthropic has been building its own first-party skills moat:

• anthropics/skills repository with curated first-party skills
• frontend-design skill at 277,000+ installs — the most-installed skill in the ecosystem
• Skills bundled directly into Claude Code (skill-creator, mcp-builder, webapp-testing, pdf, docx, pptx, xlsx)
• The /design skill that launched as a first-party Claude Code feature

The difference is tone, not trajectory. OpenAI deprecated the open repo and pointed to the managed directory. Anthropic keeps the open repo alive while shipping first-party skills that are hard to compete with because they integrate deeper than any third-party skill can. Both strategies end in the same place: the platform owner controls the defaults, and defaults win.

View on Substack →

What Happens Next

The agent skills ecosystem report from Agentman counted eight major marketplaces by Q2 2026, up from one in December 2025. That number will contract. The pattern from every prior platform cycle predicts consolidation:

1. Platform-owned directories (OpenAI Plugin Directory, Anthropic Claude Skills) will absorb the top 50-100 skills that matter.
2. Cross-vendor aggregators that index across all platforms will survive if they solve discovery better than any single vendor's search.
3. Vertical skill packs (scientific skills, finance skills, legal skills) will thrive because platform owners cannot curate domain expertise.
4. Generic directories that index one vendor's ecosystem will die or get acqui-hired.
5. Commercial marketplaces (Agensi) survive only if they crack monetization — something the platform owners have not attempted yet for skills.

The window for building a defensible skills directory is not closed. But it narrowed significantly on June 22, when OpenAI's deprecation notice turned an open catalog into a managed marketplace. The startups that saw this coming six months ago are repositioning. The ones that did not are about to learn what "platform risk" means in practice.

💡 Bottom line: The agent skills format won. It is an open standard on 40+ platforms. But the distribution layer around that format is being captured in real time. Build skills on the open standard. Publish through the first-party directories. And if you are running a third-party directory, your survival depends on being cross-vendor — the only position the platform owners cannot replicate without helping their competition.

---

Originally published at AgentConn