Browse by Category
Pick a category to jump straight to its ranked list. Every tool is filed by primary function. From MCP servers that extend AI coding agents to prompt libraries and full agent frameworks, find exactly what you need.
- MCP Server — Model Context Protocol servers for AI agents
- Claude Skill — Skills and plugins for Anthropic Claude
- Codex Skill — Skills for OpenAI Codex agents
- Agent Tool — General-purpose AI agent utilities
- Prompt Library — Curated prompt collections and templates
- AI Coding Assistant — AI-powered code generation and editing tools
- AI Tool — Other open-source AI tools
Popular Scenarios
Scenario pages answer "what's the best tool for X?" with a ranked, evidence-backed shortlist. Each scenario page ranks tools by quality score, stars, and community activity to help you find the right solution for your use case.
Why AgentSkillsHub?
Agent Skills Hub is a security-graded directory of open-source AI agent skills and MCP servers — it answers "is this safe to install?" before you run a stranger's code inside your agent. It indexes 150,000+ GitHub repositories across MCP servers, Claude skills, Codex skills, AI coding assistants, and agent tools, scores each on 10 weighted signals, and refreshes every 8 hours.
Quality Scoring
Every tool gets a 0–100 composite score across six dimensions: completeness, clarity, specificity, examples, README structure, and agent readiness. The score separates production-ready projects from experiments, so you can filter without reading 200 READMEs.
Security Grading & Sources
Every indexed repository carries a security grade — SAFE, CAUTION, UNSAFE, or UNAUDITED. Grades come from a rule-based scan built on the SlowMist agent-security taxonomy, covering 11 red-flag categories including credential harvesting, data exfiltration, and curl | sh installers.
The scale of the problem is documented in independent research. As Liu et al. (2026) report in their study of 31,132 agent skills:
"26.1% of agent skills contain security vulnerabilities."
Our own full-catalog census is published openly as a citable dataset: DOI 10.5281/zenodo.21292799 (CC-BY-4.0), mirrored on Hugging Face. The grading method is a first-layer rule-based scan, not a line-by-line human audit — its limits are documented in full on our methodology page.
Compare and Discover
Use the Compare feature to evaluate multiple tools side by side. The Skill Analyzer provides security grades and platform compatibility checks, helping teams make informed decisions about which open-source AI tools to adopt.
How It Works
Discovery is automated end to end: GitHub → scan → grade → publish, every 8 hours. Agent Skills Hub discovers tools from GitHub using multi-phase data collection. Each repository goes through search, enrichment, README analysis, and quality scoring. Our system tracks stars, forks, commit frequency, open issues, documentation quality, and community engagement to produce a single composite score. New tools appear within hours of being published. Developers can browse by category, search by keyword, filter by programming language, or explore curated scenario pages like browser automation, code review, and database management.
Stay Updated
One email a week, Mondays: the top 20 movers by star velocity. Subscribe for trending AI agent tools, new MCP servers, and top skill picks delivered to your inbox every Monday. Join thousands of developers who rely on AgentSkillsHub to stay current with the fast-moving AI agent ecosystem. Each issue features the top new tools of the week, star velocity rankings, and curated recommendations across all categories.
About the Editor
AgentSkillsHub is built and maintained by Jason Zhu (@GoSailGlobal), an independent researcher tracking open-source AI agent ecosystems since 2024. Jason is also the author of the Blue Book of Agent Skills 2026, an in-depth analysis of the Claude Skills / MCP / Codex ecosystem published openly on GitHub.
The Hub runs as a solo project, designed to be transparent and reproducible: every scoring dimension is documented, every data snapshot is archived, and the entire source code is open under MIT on GitHub. The goal is not to rank skills by popularity alone, but to help developers find tools that are actually production-ready.
Editorial Methodology
Every indexed repository passes through a 6-phase pipeline running every 8 hours: GitHub search → metadata enrichment → README fetch → quality scoring (10 weighted dimensions) → category classification → composability analysis. No manual curation is applied to ranking — signals are purely data-driven to avoid editorial bias.
Daily reports of new skills and weekly trending lists are selected according to transparent rules documented in the repository. When human judgement is needed (such as for "Verified Creator" decisions), the reasoning and criteria are written down publicly, not hidden behind a pay-for-placement system.