Three months ago, agent skills barely existed as a category. Developers had hacked together CLAUDE.md files that gave their coding agents new tricks, but there was no registry, no standard format, and no way to find what was out there.
The AgentNDX skills directory now tracks 95 curated skills across 12 categories. Small compared to the 25,000+ MCP servers in the index — but growing faster than MCP servers did at the same stage.
Here’s what we’re seeing, who’s building, and where the gaps are.
The Numbers
AgentNDX tracks 95 agent skills as of late July 2026. Every skill gets the same review process as MCP servers: verified install path, confirmed compatibility, and a health check.
The breakdown by category:
| Category | Skills | Share |
|---|---|---|
| Engineering | 20 | 21.1% |
| Tool | 10 | 10.5% |
| Data | 9 | 9.5% |
| Creative | 8 | 8.4% |
| Content | 7 | 7.4% |
| Automation | 7 | 7.4% |
| Orchestration | 6 | 6.3% |
| Memory | 6 | 6.3% |
| Compliance | 6 | 6.3% |
| Security | 5 | 5.3% |
| Product | 5 | 5.3% |
| Finance | 4 | 4.2% |
Engineering dominates. No surprise — developers building tools for themselves is the oldest pattern in software. But the second tier is more interesting: tool utilities, data, and creative skills are growing fast, which tells you people are pushing agents well past code review.
Who Builds Skills
Of the 95 tracked skills, 26 come from vendors — Anthropic, HashiCorp, Vercel, Datadog, and a handful of others. The remaining 69 are community-built.
Compare that to MCP servers, which skew heavily vendor-published. Skills are the opposite: community developers write the majority. Quality is less consistent, but you see more experimental ideas making it out the door.
Three repos punch above their weight:
- alirezarezvani/claude-skills (5,200 stars) — Research, email, SQL, product specs, and dependency auditing
- VoltAgent/awesome-agent-skills (4,800 stars) — Context management, multi-agent routing, knowledge graphs
- sickn33/antigravity-awesome-skills (22,000 stars) — Long-form writing, design, creative tools
Anthropic’s own skills repo (3,200 stars) includes the Web App Tester and MCP Server Generator — both ended up as de facto templates for how to structure a skill well.
Agent Compatibility
Not every skill works with every agent. The spread is wider than you’d expect:
| Agent | Skills Supported |
|---|---|
| Claude Code | 93 |
| Cursor | 66 |
| Universal | 60 |
| Codex | 46 |
| Windsurf | 12 |
| Gemini CLI | 4 |
Claude Code at 93 is expected — skills started there, via CLAUDE.md files. Cursor at 66 makes sense too; its instruction system maps well to the skill format.
“Universal” means any agent that reads markdown instruction files. In practice: Claude Code, Cursor, and most fork-based agents. Codex at 46 is catching up.
Windsurf and Gemini CLI are the trailing edge. Twelve and four, respectively. Both support the format on paper, but almost nobody tests against them. We’d like to see that change — two of the largest agent platforms shouldn’t be afterthoughts.
Free vs. Pro
62 free, 32 pro. The paid skills cluster in orchestration, memory, and compliance — categories where the logic is genuinely complex enough to justify a price.
If you’re kicking the tires, start with the free tier. It covers code review, writing, data queries, and deployment validation — the workflows most people reach for first.
Where the Gaps Are
The category list tells you what exists. The more interesting question is what’s missing.
Testing. Engineering has 20 skills but only one — Web App Tester — does end-to-end testing. No unit test generation, no integration scaffolding, no coverage analysis. Someone will fill these. Probably soon.
IaC beyond Terraform. HashiCorp shipped a Terraform skill. Pulumi, CDK, and Ansible have nothing. If your stack doesn’t run on Terraform, you’re out of luck.
Observability. Datadog published a skill. Grafana, PagerDuty, New Relic haven’t. MCP servers are way ahead here.
Language coverage. Most skills assume TypeScript or Python. Go, Rust, Java, C# developers are underserved. If you write skills in those languages, there’s an open lane.
Mobile. Zero skills for iOS, Android, React Native, or Flutter. Given how many developers ship mobile apps, that’s a strange gap.
What to Watch
A few things we’re watching:
Vendor adoption. The vendor skill list is short: Anthropic, HashiCorp, Datadog, Vercel, Stripe, Sentry, and a handful of others. Compare that to the MCP server vendor list, which is ten times longer. When AWS or Google Cloud ship official skills — not just MCP servers — the numbers will move fast.
Composition. Right now, skills are standalone. The next wave is skills that call other skills or coordinate with MCP servers. The Workflow Automator and Multi-Agent Router gesture in that direction, but nobody’s cracked skill-to-skill communication conventions yet. We think that’s the most interesting unsolved problem in the space.
Discovery. At 95 entries, finding the right skill is already annoying. At 500, it’ll require real search, compatibility filtering, and quality signals — which is why we built the skills directory.
So what
Agent skills are roughly where MCP servers were a year ago: small enough to track the whole space, growing fast enough that you won’t be able to for much longer. Community developers are doing most of the building. The gaps are wide open. If you’ve been looking for something to contribute to, this is it.
Full directory: agentndx.ai/skills.
FAQ
Q: What is the difference between an agent skill and an MCP server? A: An MCP server is a running service your agent connects to — it provides tools, resources, and prompts through a standardized protocol. A skill is a markdown instruction file that teaches an agent how to do a specific task. Skills run inside the agent. MCP servers run outside it. We wrote a longer comparison here.
Q: How do I install an agent skill?
A: Most install with one command. For Claude Code: gh skill install [repo]/[skill-name]. That drops the instruction file into your project. Full walkthrough here.
Q: Are agent skills safe to use? A: Skills are instruction files, not executable code. They tell your agent what to do, but the agent still applies its own judgment and uses the tools already available to it. That said, a badly written skill can instruct an agent to do destructive things — so read the instructions before you install, the same way you’d review any third-party code.