anthropics/skills (~173k stars) is Anthropic's official Agent Skills repository and the reference implementation of the agentskills.io specification. It contains 100+ skills in the SKILL.md format and serves as the canonical example of how skills should be structured and distributed. As the repo from the company behind Claude, Claude Code, and the agentskills.io spec, it carries significant weight in defining the format.
google/skills (~19k stars) provides 174+ skills focused on Google Cloud, AI, and ML products. It is a narrower, product-specific collection compared to Anthropic's general-purpose approach, but it demonstrates that major cloud providers are investing in the skills distribution model.
microsoft/skills (~3k stars) contains 174 skills for Azure SDKs. Microsoft's involvement is notable given its simultaneous consolidation of its own agent infrastructure — Semantic Kernel is in maintenance mode since April 2026, with all greenfield work redirecting to the Microsoft Agent Framework.
agentskills/agentskills (~24k stars) is the specification repository for the Agent Skills standard (agentskills.io). It defines the format, metadata schema, and progressive disclosure model that underpins the ecosystem. With 25.7k stars and 1.9k forks, it serves as the governance layer for the standard.
Engineering-Focused Skill Frameworks
mattpocock/skills (~245k stars, surging) is an engineering skills collection for real engineers, offering TDD, code review, and architecture skills. It supports dual install paths and has experienced rapid star growth, making it one of the most popular skill repos.
addyosmani/agent-skills (~49k stars) provides 23 production-grade engineering skills with auto-activation and compatibility with 70+ agents. Its focus on production readiness and broad agent compatibility distinguishes it from more specialized collections.
alexander-kastil/skills-collection (~49k stars) offers 860+ universal agentic skills across 8 platforms, making it one of the broadest cross-platform skill collections available.
alirezarezvani/claude-skills (~25k stars) contains 380 Claude Code skills spanning multiple domains. It is a focused, Claude-specific collection with detailed documentation.
Multi-Agent Orchestration Platforms with Skills Systems
crewAIInc/crewAI (~58k stars, 8.6k forks) provides multi-agent orchestration with a built-in skills system using SKILL.md packages and progressive disclosure. Its skills system is tightly integrated with the broader crew architecture, making it a natural choice for teams building multi-agent workflows.
Significant-Gravitas/AutoGPT (~187k stars) is the pioneering autonomous agent project that popularized the concept of breaking tasks into subtasks and executing them autonomously. It includes a visual builder and plugin ecosystem. However, the original AutoGPT repo now returns a 404 (deprecated), and the project's active development has shifted elsewhere.
langchain-ai/langchain (~146k stars, 24.6k forks) is the dominant agent engineering platform. Skills are distributed through LangSmith, and the framework reached 1.0 in September 2026, signaling maturity past its experimental phase.
Significant-Gravitas/AutoGPT (~187k stars) is listed as a key repo but should be noted as effectively deprecated — the original repo returns 404. Its historical importance in pioneering autonomous agents remains significant.
Notable Specialized Projects
affaan-m/ECC (~242k stars) is described as an "Agent Harness Operating System" with 261 skills, 64 agents, and 84 commands. It operates as a cross-harness OS with built-in security scanning, positioning it as infrastructure rather than just a skills collection.
JuliusBrussee/caveman (~108k stars) went viral as a skill-plus-proxy for coding agents, demonstrating a 65% token reduction through its "caveman speak" skill. It represents the kind of viral, single-purpose skill innovation that drives ecosystem growth.
vercel/eve (~5k stars) is a filesystem-first agent framework that loads skills from an agent/skills/ directory. Its approach is minimal and developer-friendly, aligning with Vercel's broader philosophy of simple, deployable tooling.
millionco/react-doctor (~15k stars) generates installable skills from codebase audits, bridging the gap between static analysis and agent skill creation.
2.2 The Agent Skills Format
The Agent Skills format has emerged as the universal packaging convention for agent capabilities. A skill is a folder containing:
- SKILL.md — the core file with YAML frontmatter (metadata) followed by Markdown instructions
- scripts/ — executable scripts that the agent can run on demand
- references/ — reference materials and documentation
- assets/ — static assets like images, templates, or configuration files
The progressive disclosure model is central to how skills work: agents load a skill's name and description at startup, fetch full instructions only on activation, and load resources and scripts only when needed. This minimizes token usage and startup overhead while keeping capabilities available on demand.
The agentskills.io open standard, maintained in the agentskills/agentskills spec repo, is emerging as the universal format. Anthropic's spec is the de facto standard, adopted by Claude, Claude Code, Cursor, Windsurf, and Kiro. The SKILL.md format (YAML frontmatter + Markdown instructions) is now supported across 70+ agents, according to addyosmani/agent-skills.
The Agent Plugins 1.0.0 standard (August 6, 2026) aims to formalize skill distribution further, but adoption is still early and compatibility across platforms is not guaranteed.
2.3 Skill Frameworks and Tooling
Microsoft Agent Framework (13.8k stars, 2.4k forks) provides a comprehensive skills architecture with file-based, inline, and class-based skill types, plus MCP integration. It is the official successor to both Semantic Kernel (maintenance mode since April 2026) and AutoGen. The framework supports YAML declarative workflow definitions, checkpointing, and HITL (human-in-the-loop) via RequestPort.
crewAIInc/skills (42 stars) is the official skills companion to the crewAI orchestration framework. It is small but represents the tight integration between orchestration and skills packaging.
skills.rest (754,000+ curated skills) is the largest known skills marketplace. It serves as a curated catalog where developers can discover and adopt skills across platforms.
skills.sh is a Claude Code plugin marketplace specifically for distributing skills. It is smaller (~2,000 skills) but focused on the Claude ecosystem, which gives it a clear niche.
MCP (Model Context Protocol) is becoming the standard interoperability layer for agent-to-skill communication. The MCP Registry lists approximately 25,125 active servers and serves as the backbone for tool-integration skills. However, security concerns are significant: 23% of entries lack a source repository, and 543 of approximately 50,000 manifests violate schema validation.
IETF ZTDS (Zero-Trust Data Sanitization) is emerging as a standard skill category for PII/PHI handling. Security-focused skills are a growing category, with ZTDS-related PRs appearing on crewAI and LangChain repositories.
2.4 Marketplaces and Registries
The skill marketplace ecosystem is fragmented but growing rapidly. The table below summarizes the major marketplaces:
| Marketplace | Approximate Size | Notes |
|---|---|---|
| Skills Directory | 635,000+ skills | Largest known collection; broad coverage |
| ClaudeSkills.com | 191,000+ skills | Focused on Claude-compatible skills |
| LobeHub | 169,000+ skills | Multi-platform skill marketplace |
| AgentArea | 138,000+ skills | Community-driven skill exchange |
| SkillsMP | 3,000,000+ skills | Largest by raw count; likely includes duplicates/test entries |
| skillmarketplace.ai | 400,000+ skills | Commercial marketplace |
| Agensi | 200+ skills | Curated, higher-quality collection |
| Skills.sh | ~2,000 skills | Smaller, focused marketplace |
| Claude Skills Hub | 42,000+ skills | Hub for Claude-specific skills |
| Agent Skill Exchange | 2,500+ skills | Exchange-oriented with community review |
The wide variation in marketplace sizes — from 200 to over 3 million — reflects differences in curation standards, deduplication practices, and whether the count includes test or draft entries. SkillsMP's 3 million figure is an outlier and likely includes many low-quality or duplicate entries. The more established marketplaces (ClaudeSkills.com at 191K, LobeHub at 169K) tend to have better quality control.
2.5 Community and Ecosystem Trends
The ecosystem is shifting from monolithic frameworks to modular skill systems. This is the most significant structural trend: developers increasingly prefer composing discrete, reusable skills over building monolithic agent applications.
Fork counts reveal the community's engagement patterns:
langchain-ai/langchain: 24.6k forksanthropics/skills: 21k forkscrewAIInc/crewAI: 8.6k forks
These high fork counts indicate active forking for customization and integration, which is a sign of a healthy, developer-driven ecosystem.
AI-generated contributions are common on major repos, reflecting the growing use of AI coding agents to contribute to AI infrastructure projects. This creates a feedback loop where AI agents help build the tools that AI agents use.
Microsoft's consolidation is a notable strategic move: Semantic Kernel is in maintenance mode, AutoGen is effectively deprecated, and all greenfield work flows into the Microsoft Agent Framework. This mirrors the broader industry trend of consolidation around fewer, more capable platforms.
AutoGPT is effectively deprecated — the original repo returns a 404 error. While it pioneered the autonomous agent concept, it has not maintained its development momentum.
Community-driven collections play an important curation role. The most prominent include:
VoltAgent/awesome-agent-skills(35,000 stars, 1,497+ skills)skills-library/agentic-skills(41,300 stars, 1,600+ skills)multica-ai/andrej-karpathy-skills(156,000 stars)alirezarezvani/claude-skills(18,700 stars, 337 skills)ComposioHQ/awesome-claude-skills(~16,000 stars, 1,000+ skills)
The "awesome" pattern — borrowing from the awesome-* lists popularized in the open-source world — provides a familiar navigation model for developers in an ecosystem where the sheer volume of repositories makes discovery difficult.
2.6 Security and Emerging Standards
Security is a top concern in the agent skills ecosystem. Key issues include:
- MCP Registry supply-chain risk: 23% of entries lack a source repository, and 543 of ~50,000 manifests violate schema validation. Agents that dynamically load MCP servers from the registry could execute unverified code, creating a supply-chain vulnerability analogous to risks seen in the npm and PyPI ecosystems.
- IETF ZTDS (Zero-Trust Data Sanitization): Emerging as a standard skill category for PII/PHI handling. ZTDS-related PRs are appearing on crewAI and LangChain, indicating that security is becoming a first-class concern in skill development.
- AI-generated contributions: Common on major repos, which raises questions about the reliability and security review of contributions that are not human-authored.
The Agent Plugins 1.0.0 standard (August 6, 2026) is the first attempt to formalize skill distribution with security and compatibility in mind, but adoption is still early.
Comparison Table
The following table compares the top 15 repositories across key dimensions:
| Repository | Stars | Forks | Primary Role | Skills Count | Format | Status |
|---|---|---|---|---|---|---|
| anthropics/skills | ~173k | 21k | Official reference impl | 100+ | SKILL.md | Active |
| mattpocock/skills | ~245k | — | Engineering skills | — | SKILL.md | Active, surging |
| affaan-m/ECC | ~242k | — | Agent harness OS | 261 | Custom | Active |
| Significant-Gravitas/AutoGPT | ~187k | — | Autonomous agent | — | Custom | Deprecated (404) |
| langchain-ai/langchain | ~146k | 24.6k | Agent platform | Via LangSmith | Custom | Active, 1.0 |
| JuliusBrussee/caveman | ~108k | — | Coding agent skill+proxy | 1 (viral) | Custom | Active |
| addyosmani/agent-skills | ~49k | — | Production engineering skills | 23 | SKILL.md | Active |
| alexander-kastil/skills-collection | ~49k | — | Universal skills | 860+ | Multi-platform | Active |
| crewAIInc/crewAI | ~58k | 8.6k | Multi-agent orchestration | Built-in | SKILL.md | Active |
| alirezarezvani/claude-skills | ~25k | — | Claude-specific skills | 380 | Custom | Active |
| agentskills/agentskills | ~24k | 1.9k | Specification repo | — | Spec | Active |
| microsoft/semantic-kernel | ~29k | — | Enterprise SDK | — | Custom | Maintenance mode (since Apr 2026) |
| google/skills | ~19k | — | Google Cloud/AI skills | 174+ | SKILL.md | Active |
| vercel/eve | ~5k | — | Filesystem-first agent framework | From agent/skills/ | Custom | Active |
| millionco/react-doctor | ~15k | — | Codebase audit → skills | Generated | SKILL.md | Active |
Trend Analysis
From Monolithic Frameworks to Modular Skill Systems
The single most important structural trend is the shift away from monolithic agent frameworks toward modular, composable skill systems. Five years ago, the dominant pattern was a single framework (LangChain, AutoGPT) that provided everything an agent needed. Today, the trend is toward small, focused skill packages that can be mixed and matched across different agent runtimes. This mirrors the broader software industry's shift from monolithic applications to microservices and package ecosystems.
The evidence for this trend is strong: the proliferation of "awesome" collections, the growth of dedicated skill marketplaces, and the emergence of the agentskills.io standard all point toward a modular future. Even large frameworks like LangChain and crewAI are adding skills packaging on top of their existing architectures rather than replacing them.
Anthropic's Spec as the De Facto Standard
Anthropic's agentskills.io specification has become the de facto standard for agent skills, adopted by Claude, Claude Code, Cursor, Windsurf, and Kiro. This is significant because it means a single company's specification is shaping how the entire ecosystem packages and distributes agent capabilities. The spec's progressive disclosure model (name/description at startup, full instructions on activation, resources/scripts on demand) has proven to be an effective pattern that balances discoverability with efficiency.
The Rise of Marketplaces
Skill marketplaces are emerging as a distinct layer in the ecosystem, separate from both the spec/standard layer and the framework layer. The MCP Registry (~25,125 active servers) and dedicated marketplaces like Skills Directory (635,000+ skills) and ClaudeSkills.com (191,000+ skills) are becoming the places where developers discover and adopt skills. However, quality control remains a concern — the wide variation in marketplace sizes and the lack of standardized quality metrics mean that developers must vet skills individually.
Security as a First-Class Concern
Security is no longer an afterthought in the agent skills ecosystem. The MCP Registry's supply-chain vulnerabilities (23% of entries without source repos, 543 schema violations), the emergence of IETF ZTDS as a standard skill category, and the growing recognition of AI-generated contributions as a security concern all indicate that the ecosystem is grappling with its trust model. The Agent Plugins 1.0.0 standard (August 2026) is an early attempt to address these concerns through formalized distribution.
Microsoft's Consolidation
Microsoft's consolidation of its agent infrastructure — retiring Semantic Kernel and AutoGen in favor of the Microsoft Agent Framework — is a strategic move that mirrors the broader industry trend of consolidation. The Agent Framework (13.8k stars, 2.4k forks) is positioned as the unified platform for enterprise agent development, with skills architecture built in from the ground up. This consolidation may reduce fragmentation in the Microsoft ecosystem but increases vendor lock-in risk.
Recommendations
For Developers Building Agent Skills
Adopt the SKILL.md format with agentskills.io metadata. The progressive disclosure model is well-supported across 70+ agents and is the most portable format available. Avoid proprietary formats unless you have a specific reason to lock into a single runtime.
Use community collections for discovery. The "awesome" collections (VoltAgent/awesome-agent-skills, skills-library/agentic-skills) provide curated entry points into the ecosystem. Start with these rather than browsing marketplaces directly.
Validate skills against the spec. The
agentskills/agentskillsspec repo provides the authoritative schema. Validate your SKILL.md files against it before publishing to ensure compatibility.
For Teams Evaluating Frameworks
Prefer frameworks with native skills support. CrewAI, LangChain, and the Microsoft Agent Framework all have built-in skills packaging. Choosing a framework that natively supports the SKILL.md format reduces integration friction.
Monitor the MCP Registry security posture. Before dynamically loading MCP servers from the registry, verify that the entry has a source repository and a valid manifest. The current 23% missing-source rate is a material risk for production deployments.
For Organizations Investing in Agent Infrastructure
Plan for format fragmentation. While SKILL.md is the leading format, the ecosystem is not yet fully standardized. The Agent Plugins 1.0.0 standard (August 2026) may resolve this, but adoption is early. Build abstraction layers that can accommodate format changes.
Invest in security tooling for skills. The supply-chain risks in the MCP Registry and the prevalence of AI-generated contributions mean that security review of skills should be a first-class process, not an afterthought. Consider implementing IETF ZTDS-compliant sanitization for skills that handle PII or PHI.
Track the Microsoft Agent Framework consolidation. If you are in the Microsoft ecosystem, plan your migration from Semantic Kernel and AutoGen to the Microsoft Agent Framework. The maintenance-mode status of the older tools means they will not receive new features or security patches.