Skills.sh: Vercel’s Open Agent Skills Directory

Source: skills.sh | Author: Vercel Labs | Published: 2026-01-20 Category: product-announcement | Credibility: medium

Executive Summary

• Skills.sh is Vercel’s directory and leaderboard for AI agent skills — reusable SKILL.md-based instruction packages installable via npx skills add <owner/repo> across 40+ AI coding agents (Claude Code, Cursor, GitHub Copilot, Gemini CLI, and many more). The site reports 91,000+ total skill installations and has indexed 87,000+ unique skills since its January 2026 launch.
• The platform surfaces a genuine ecosystem need — centralizing discovery of Agent Skills Specification-compatible packages — but suffers from significant quality and security problems. An independent audit found 12% of 2,857 audited skills were malicious, and community feedback consistently reports the majority of listings are low-quality AI-generated content.
• Skills.sh is a directory, not a runtime or specification. It sits atop the open Agent Skills Specification (agentskills.io) and the open-source skills CLI (github.com/vercel-labs/skills, 13.1k stars). The underlying spec has genuine cross-industry adoption (Anthropic, Microsoft, Google, OpenAI), making it the de facto packaging standard for AI agent procedural knowledge.

Critical Analysis

Claim: “91,000+ total skill installations across the ecosystem”

• Evidence quality: vendor-sponsored
• Assessment: The installation numbers are self-reported by Vercel and not independently verifiable. The “find-skills” skill alone accounts for 774.9K installs, which is the meta-skill for discovering other skills — suggesting significant inflation from bootstrapping behavior. Microsoft Azure skills reportedly total 2.3M installs across multiple repositories, but these numbers come from the platform itself with no third-party audit of counting methodology.
• Counter-argument: Install counts on any package registry (npm, PyPI) are often inflated by CI pipelines, bots, and automated tooling. The top skills are overwhelmingly from verified vendors (Vercel, Anthropic, Microsoft), which suggests organic adoption in their respective ecosystems is real, even if raw numbers are unreliable as a quality signal.
• References:
  • InfoQ: Vercel Introduces Skills.sh — independent coverage noting “tens of thousands of installs shortly after launch”
  • Vibecoding: Skills.sh Review (2026) — independent review scoring 3.5/5 with infrastructure 5/5 but execution 2/5

Claim: “Open ecosystem — reusable capabilities for AI agents”

• Evidence quality: vendor-sponsored (with partial independent corroboration)
• Assessment: The “open” framing is partially accurate. The underlying Agent Skills Specification is genuinely open (Apache-2.0, published at agentskills.io, with contributions from Anthropic, Microsoft, Google). The skills CLI is open-source (13.1k GitHub stars, 1.1k forks). However, skills.sh the directory is a Vercel product — the data aggregation, ranking algorithm, install tracking, and discovery experience are controlled by Vercel. This mirrors the npm/GitHub pattern: open format, vendor-controlled marketplace.
• Counter-argument: Alternative directories already exist (Skills Directory, agentskill.sh, ClawHub, OpenAI’s Codex skills), suggesting the spec is genuinely portable even if skills.sh is the largest hub. The real lock-in risk is low because skills are just markdown files in GitHub repos — the directory adds discovery, not dependency.
• References:
  • KDnuggets: Top 5 Agent Skill Marketplaces — lists multiple competing directories
  • GitHub: vercel-labs/skills — open-source CLI repository (13.1k stars)

Claim: “Works with 40+ AI agent platforms”

• Evidence quality: vendor-sponsored (with independent corroboration)
• Assessment: Cross-agent compatibility is real and independently verified. Microsoft documents Agent Skills support in VS Code/GitHub Copilot, Google’s Gemini CLI supports them, and numerous open-source agents (OpenCode, Pi Coding Agent, Goose) implement the specification. The skills CLI detects installed agents and symlinks skills to agent-specific directories. The shared format is the Agent Skills Specification’s SKILL.md, which is genuinely interoperable.
• Counter-argument: “Works with” is generous — not all agents implement the spec identically. Some load skills as static context, others support progressive disclosure, and a few support the allowed-tools field for pre-approved tool execution. The experience is not uniform. Testing by the vibecoding.app review confirmed skills worked across Claude Code, Cursor, and GitHub Copilot but noted behavioral differences.
• References:
  • Microsoft: Use Agent Skills in VS Code — Microsoft’s official adoption
  • LlamaIndex: Skills vs MCP tools for agents — independent comparison

Claim: “Automated security audits now available for skills.sh” (implied by Snyk/Socket partnerships)

• Evidence quality: vendor-sponsored
• Assessment: Vercel has partnered with Snyk and Socket to add supply-chain security scanning to skills.sh. Every skill installation triggers automated security analysis, and malicious skills are flagged/blocked. This is a meaningful step, but the Grith.ai/Koi Security audit finding 12% malicious skills across 2,857 audited skills (341 malicious) demonstrates the problem is severe. The attack surface is unique: skills combine natural-language prompt instructions with executable scripts, creating vectors for prompt injection, silent data egress, and CI pipeline compromise that traditional code scanning tools may not catch.
• Counter-argument: Security scanning partnerships are a response to a real problem, not proof the problem is solved. The 12% malicious rate is alarming and suggests the ecosystem grew faster than security infrastructure. The install-count ranking system incentivizes quantity over quality, and there is no formal review or certification process before listing. Enterprise teams should treat skills like untrusted npm packages.
• References:
  • Grith.ai: We Audited 2,857 Agent Skills. 12% Were Malicious. — independent security audit
  • Snyk: Securing the Agent Skill Ecosystem — Snyk’s partnership announcement
  • Socket: Supply Chain Security for skills.sh — Socket’s partnership announcement

Credibility Assessment

• Author background: skills.sh is created by Vercel Labs, the innovation arm of Vercel (the company behind Next.js, Turborepo, and the Vercel platform). Vercel is a well-funded company ($563M+ raised) with strong credibility in the frontend/developer tools space.
• Publication bias: This is a vendor product page. All claims, statistics, and positioning are self-serving. The install counts, ecosystem breadth claims, and “open” framing should be evaluated against independent evidence.
• Verdict: medium — The product is real and the underlying specification has genuine cross-industry adoption. However, the directory itself is a Vercel marketing and ecosystem play. Critical quality and security problems documented by independent reviewers significantly undermine the “production-ready” positioning. Trust vendor-published skills; be deeply skeptical of community-submitted skills.

Entities Extracted

Entity	Type	Catalog Entry
Skills.sh	vendor	link
Agent Skills Specification	open-source	link
Agent Skill Supply Chain Risk	pattern	link