Published: July 30, 2026 | Reading Time: ~18 minutes | Channel: techminute
If you opened GitHub trending this morning, the #1 repository isn't a new framework, a leaked model, or a crypto whitepaper. It's a skill. A SKILL.md file — 64 lines of Markdown plus some Python scaffolding — with 55,294 stars, 377 earned today alone, and an installation command that works across 20 different AI coding agents without a single modification.
The #2 repo, with 18,406 stars and 916 earned today, is an open-source desktop app for sharing those same skills.
Nine months ago, "agent skills" didn't exist as a category. Today, it's an ecosystem with roughly 1.9 million public skills indexed across eight marketplaces, supported by 40-plus platforms from Claude Code to GitHub Copilot to VS Code, with at least one marketplace paying creators an 80/20 revenue split and another backed by Vercel. The standard — a simple YAML frontmatter + Markdown body — is stewarded by the Agentic AI Foundation, the same body that governs MCP. And researchers just published the first peer-reviewed benchmark proving that well-built skills raise agent pass rates by an average of 16.2 percentage points, with healthcare alone gaining 51.9 points.
The "app store for AI agents" is not a metaphor. It's happening. And like every gold rush, most of what's being dug up isn't worth much — but some of it is worth everything.
To understand how we got here, you have to appreciate the speed. Anthropic introduced Agent Skills as a Claude Code feature in October 2025. On December 18, 2025, it published the specification as an open standard at agentskills.io and handed stewardship to the Agentic AI Foundation — the same neutral body that governs the Model Context Protocol. The parallel is deliberate: MCP gave agents the ability to connect to tools. Skills give them the ability to learn procedures. Together, they're the hands and the judgment.
The format is almost absurdly simple. A skill is a directory containing a SKILL.md file with YAML frontmatter:
---
name: pdf-processing
description: Extracts text and tables from PDF files, fills PDF forms, and merges multiple PDFs. Use when working with PDF documents.
license: Apache-2.0
---
# Instructions
Step-by-step workflow for PDF handling...
That's it. Name (1-64 chars), description (1-1024 chars), and Markdown instructions. Optional directories for scripts, references, and assets. The magic is in progressive disclosure: an agent pre-loads only the ~100-token name and description for all installed skills, then loads the full instructions only when a task matches. A developer can have 200 skills installed and pay zero context cost until one actually fires.
What happened next was not planned by Anthropic — or at least, not exclusively. OpenAI had quietly shipped a structurally identical implementation in Codex and ChatGPT, using the same file naming, metadata format, and directory layout. VentureBeat documented this "convergent evolution" in late 2025. When two competing frontier labs independently land on the same format, you're no longer looking at a feature. You're looking at infrastructure.
By January 2026, Skills.sh — backed by Vercel — launched an npm-style CLI: npx skills add. Claude Code 2.1.0 added hot-reloading so skills update mid-session without a restart. By June 2026, approximately 40 products appeared on the agentskills.io showcase, including GitHub Copilot, Cursor, VS Code, Gemini CLI, Databricks Genie Code, Snowflake Cortex Code, and JetBrains Junie.
A skill written once now runs on Claude Code, Codex, Cursor, Copilot, Gemini CLI, and a dozen other tools without modification. Portability is the whole game.
Let's take stock of what nine months of open-standard adoption looks like, because the numbers tell a story that the headlines miss.
The Catalogs. SkillsMP indexes approximately 1.9 million public skills scraped from GitHub repositories. Skills.sh lists hundreds of thousands across community submissions. LobeHub aggregates 169,000-plus skills. There are now over 2,500 Claude Code plugin marketplaces registered at claudemarketplaces.com — ranging from polished collections by teams like Trail of Bits and Vercel Labs to abandoned single-skill repos. Eight major marketplaces operate as of Q2 2026, up from exactly one (Anthropic's official directory) in December 2025.
The Categories. Skills.sh's category index, one of the largest open catalogs, breaks down the landscape:
| Category | Published Skills |
|---|---|
| Development & Engineering | 288,811 |
| Product Management | 86,948 |
| Marketing | 74,510 |
| Data & Analytics | 69,187 |
| Operations | 51,007 |
| Sales | 42,570 |
| Design | 25,743 |
| Legal | 17,624 |
| Finance & Accounting | 14,932 |
| Healthcare & Life Sciences | 6,354 |
Software engineering dominates by raw volume — no surprise, since developers were the first audience and models already have strong coding coverage. But the fastest-growing long tail is in business functions: product, marketing, operations, and regulated verticals like legal and healthcare. These are the domains where skills deliver the largest measured performance gains, even though they have the smallest catalog counts. Volume and value point in opposite directions.
The Marketplaces. Not all directories are created equal. The table that matters:
| Marketplace | Catalog | Curation | Security Review | Creator Payments |
|---|---|---|---|---|
| SkillsMP | ~1.9M | GitHub scrape, 2+ stars | None | No |
| Skills.sh | Hundreds of thousands | Community, npm-style | None | No |
| LobeHub | 169K+ | Scraped | None | No |
| ClaudeSkills.info | 658 | Community + official | None | No |
| Anthropic official | ~20 | Manual, Anthropic-verified | Internal | No |
| Agensi | 200+ | Manual + automated | 8-point scan | Yes, 80/20 split |
The catalog size gap between SkillsMP's 1.9 million and Agensi's 200-plus is not a bug — it's the difference between scraping and curation. And as we're about to see, most of those 1.9 million skills shouldn't be installed on anything you care about.
The single most important finding in this ecosystem dropped recently from SkillsBench — a peer-reviewed benchmark built by researchers from Stanford, CMU, Berkeley, Oxford, and BenchFlow. They tested 84 tasks across 11 domains using 7,308 trajectories, and measured how agent skills affect performance.
Curated skills raised pass rates by an average of 16.2 percentage points. Healthcare, the domain with the largest gain, jumped 51.9 points. The more organizational expertise a domain requires, the more skills matter.
But here's the catch. The same benchmark analyzed 47,150 publicly available skills and found an average quality score of 6.2 out of 12. The researchers only used top-quartile skills — those scoring 9 or above — in their performance tests. Focused, modular skills (2-3 per task) delivered +18.6 points, while monolithic "everything in one file" skills reduced performance by 2.9 points.
The practical translation: if you grab a random skill off a scraped catalog like SkillsMP, there's roughly a 75% chance it won't clear the quality bar that produces measurable improvements. Skills work — dramatically so — but only when they're well-built, focused, and reviewed.
If the quality gap is disappointing, the security picture is alarming. An audit of 22,511 skills across four sources — skills.sh, ClawHub, GitHub, and Tessl — surfaced 140,963 issues, averaging 6.3 issues per skill. Some were benign (formatting errors, missing optional fields). Many were not.
Snyk's ToxicSkills research, published alongside the Agensi security program, tested skills from public catalogs and found prompt injection vulnerabilities in 36% of skills tested. Thirty-six percent. Think about that number. More than one in three publicly available skills contains the AI equivalent of an unsanitized SQL query — instructions that could be exploited to override the agent's behavior, exfiltrate data, or execute unintended commands.
This is not a theoretical concern. Agent skills can include executable scripts in their scripts/ directory. An unaudited skill is an unaudited dependency with code-execution rights. The practical advice from every credible source is the same: treat community skills exactly like you treat open-source packages. Read the source before installing. If a marketplace hasn't scanned it, the burden is on you.
The market is responding. Agensi introduced an 8-point security scan (prompt injection, data exfiltration, secret detection, dangerous commands, obfuscation, external fetches, credential access, privilege escalation) on every listing before it goes live. Expect this to become table stakes by Q4 2026 — directories that can't certify safety will lose enterprise buyers to ones that can.
Let me tell you about the skill sitting at #1 on GitHub trending right now, because it's the perfect case study for why this ecosystem matters beyond the raw numbers.
last30days-skill by mvanhorn has 55,294 stars, 4,773 forks, 1,150 commits, and 52 community contributors across 175 merged PRs. It's currently at version 3.11.1. It installs with a single command — /plugin marketplace add mvanhorn/last30days-skill on Claude Code, npx skills add mvanhorn/last30days-skill -g on everything else — and works across Claude Code, Codex, Cursor, Copilot, Gemini CLI, and 50-plus other platforms.
What does it do? It searches 18 sources simultaneously — Reddit, X, YouTube, TikTok, Instagram, Hacker News, Polymarket, GitHub, Digg, arXiv, Techmeme, LinkedIn, StockTwits, Threads, Pinterest, Bluesky, Perplexity, and the open web — scores results by what real people actually engaged with (upvotes, likes, views, real-money odds), and synthesizes everything into a single grounded brief.
The creator built it because they needed to keep up with AI news. "Everything changes every day and the Reddit and X nerds are always on top of it first," the README explains. "I needed better prompts, and the training data was always months behind what the community had already figured out." But it became something bigger: people use it before sales calls, meetings, trips, and hiring decisions. "If you're meeting with a CEO, have you read all their tweets and YouTube transcripts from the last 30 days? I have."
This is what a skill looks like at scale: not a trivial macro, but a research engine with 18 API integrations, real-time scoring, cross-source synthesis, and a user base larger than most SaaS startups. And it runs identically across every major AI coding agent.
The #2 trending repo today is different-ai/openwork — 18,406 stars, 4,086 commits, 1,878 forks. It's a free, open-source desktop app for sharing AI agent workflows, positioned as an open-source alternative to Claude Cowork and Codex. It runs on macOS, Windows, and Linux.
The architecture is clever. Install OpenWork once, and your skills, MCP connections, Google Workspace, and Microsoft 365 integrations become available to any compatible agent through a single MCP server. The server exposes two tools: search_capabilities finds what you can use, and execute_capability runs it. Team management, access controls, and marketplace publishing come through OpenWork Den, the organizational control plane.
OpenWork matters because it signals where the ecosystem is heading: not just individual skills, but entire workspaces and workflows as portable packages. The desktop app is optional — you can use OpenWork entirely from your existing agent — but the vision is a unified skill layer that spans agents, teammates, and machines.
Between last30days-skill and OpenWork, the two most popular repositories on GitHub today are not competing. They're complementary layers of the same stack: one is the most sophisticated research skill in the ecosystem; the other is the infrastructure for sharing everything.
The indie marketplace story is exciting, but the enterprise story is where the money is. Anthropic added organization-wide skill management for Team and Enterprise plans, letting admins provision and default-enable skills centrally. They shipped stock plug-ins for finance, legal, and HR in February 2026. Microsoft's Agent Framework 1.0, released this year, supports agentskills.io-compatible skills natively, with MCP-native tooling and A2A protocol support.
The broader signal: 68% of production agent deployments have adopted MCP or an equivalent standardized tool layer, according to a Digital Applied enterprise survey from 2026. Skills sit naturally on top of that foundation. Enterprises are building governed, internal skill libraries — not publishing to public catalogs, but using the same standard to encode domain expertise that internal teams can discover and reuse.
Salesforce unified its AppExchange, Slack Marketplace, and Agentforce ecosystems into a single agent-skill storefront by April 2026 with over 200 launch partners. Google's Agent Development Kit (ADK) has emerged as a major framework with an expanding integrations ecosystem. The agent skills story is not just an indie developer phenomenon — it's becoming the default distribution model for enterprise AI capabilities.
The agent skills ecosystem is at an inflection point. Distribution and portability are largely solved problems. The next six months will be about trust.
Security scanning becomes table stakes. After ToxicSkills found prompt injection in 36% of skills, directories that can't certify safety will lose enterprise buyers. Verification, not volume, will become the selling point.
Quality benchmarks drive procurement. SkillsBench is the first. Expect buyers to demand quality scores the way they demand security scores, and curated registries to advertise them.
MCP-based access replaces downloads. Most marketplaces still work like npm: install, store on disk, update manually. Agensi Pro is currently the only marketplace running MCP-based access end-to-end — your agent connects once, searches the full live catalog, loads skills on demand, and automatically discovers new ones. "This becomes the default within 12 months," the Agensi team predicts. "Download-based distribution still works, but the operational overhead gets heavy once you pass 20 or 30 installed skills. MCP-based access scales to thousands without any local footprint."
Consolidation is coming. Eight major marketplaces is too many for a category this size. Expect acquisitions, shutdowns, and cross-listings to reduce the active set to three or four dominant platforms.
Business-function skills overtake coding skills. Developer skills lead today, but product, marketing, operations, and regulated-vertical skills are growing faster and carry higher measured value per skill. The durable money is in deep, domain-specific procedures that encode how a particular kind of business actually works — not another JavaScript testing skill.
Spec governance will be contested. As an Anthropic-originated standard, the agentskills.io specification will face the same neutrality scrutiny MCP did. The Agentic AI Foundation's stewardship helps, but when the two largest AI companies (Anthropic and OpenAI) both ship near-identical implementations, governance isn't just bureaucratic theater — it determines whether the standard stays truly open or becomes a battleground.
The agent skills ecosystem is not a prediction anymore. It's 1.9 million published skills, 40 compatible platforms, eight marketplaces, and a peer-reviewed benchmark proving that good skills make agents 16.2 percentage points more effective. The #1 and #2 repositories on GitHub today are both agent-skill related. The "app store for AI agents" analogy is not aspirational — it's descriptive, and it's been operating since December 2025.
But the gold rush metaphor cuts both ways. Most of the 1.9 million skills are mediocre. Thirty-six percent have security vulnerabilities. Curation, not catalog size, is the differentiator that actually moves agent performance. If you're building in this space, the playbook is clear: review before installing, focus on quality over quantity, and treat community skills like any open-source dependency — with respect, and with caution.
The standard won. The marketplaces are here. The benchmark proves it works. Now the hard part: making it safe.
All claims verified against Gold-tier (official specifications, GitHub repositories, peer-reviewed benchmark data) and Silver-tier (industry analysis, marketplace documentation) sources. Each source URL was scraped and confirmed accessible. Last verified: July 30, 2026.