Agent Skills Marketplaces Compared: SkillsMP, skills.sh, Claude and OpenAI

Updated October 23, 2026: rewrote the marketplace sections with current sizes, install commands and security checks for SkillsMP, skills.sh, Anthropic’s Claude Code marketplaces, OpenAI’s plugin directory and GitHub’s gh skill, added a comparison table, and replaced the security advice with 2026 research.

Five places matter if you want Agent Skills (folders with a SKILL.md file) in October 2026. SkillsMP is the biggest search index by its own count, with more than 3 million SKILL.md files crawled from GitHub, but it installs nothing and checks nothing. Vercel’s skills.sh is the closest thing to npm for skills: an open registry of 600,000+ skills, the npx skills installer, and security audits from Gen, Socket and Snyk. Anthropic distributes skills inside Claude Code plugins through its built-in claude-plugins-official marketplace. OpenAI runs one plugin directory shared by ChatGPT and Codex. GitHub’s gh skill installs from any repository and pins versions. The file format is the same everywhere. What differs is who curates, how you install, and how much checking happens before a stranger’s instructions reach an agent with shell access.

The short version: use npx skills or gh skill for cross-agent installs, Anthropic’s official marketplace when you live in Claude Code, and SkillsMP strictly as a search engine whose results you read before you trust them.

Agent Skills Marketplaces Compared (October 2026)

What Agent Skills Actually Are (And What They Are Not)

A skill is a folder containing a SKILL.md file: YAML frontmatter with a name and a description, Markdown instructions, and optional scripts/, references/ and assets/ directories. No compiled binaries, no runtime. The full specification, and how it spread to dozens of tools, is covered in our Agent Skills open standard explainer, so here is just the shape:

---
name: database-migration
description: Writes safe, reversible database migrations. Use when the user asks to create, modify or roll back a migration.
---

# Database Migration Skill

## Instructions
1. Always check the current migration state first...
2. Generate reversible migrations by default...

At startup an agent loads only each skill’s name and description, then reads the full instructions when a task matches. You can also call a skill directly: /skill-name in Claude Code, $skill-name in Codex. For skills with side effects, like a deploy workflow, Claude Code’s disable-model-invocation: true makes the skill run only when you ask for it.

This makes skills different from MCP servers. MCP provides structured API access to external tools and data sources. Skills provide instructions and workflows, teaching the agent a way of working. You might use an MCP server to connect to your database and a skill to teach the agent your team’s migration conventions.

Marketplaces distribute skills in two shapes, and the difference matters for security. skills.sh, SkillsMP and gh skill deal in bare skill folders: text the agent reads, plus whatever scripts it may later decide to run. Claude Code, Codex and ChatGPT increasingly distribute plugins, which bundle skills with hooks, MCP servers and commands. In Anthropic’s own words, a Claude Code plugin “can execute arbitrary code on your machine with your user privileges.”

SkillsMP: The Biggest Index, Not a Store

SkillsMP (Skills Marketplace) is a search engine for SKILL.md files. It crawls public GitHub repositories, and its homepage now claims more than 3,000,000 skills across 12 domains and 50+ categories, up from about 160,000 when this post first ran in February. You can search by keyword, category or occupation and preview a skill before going to its repository. It is an independent community project, not affiliated with Anthropic or OpenAI.

Three things to know before you rely on it:

• It installs nothing. The SkillsMP docs send you to the skill’s source repository and your client’s own install instructions. In practice you copy the folder, or point npx skills add or gh skill install at the repo.
• It vets nothing. SkillsMP says a listing is not a quality certification and that it has not tested every listed skill. Its own line: “Neither a listing nor a high star count guarantees safety.”
• The headline number is a file count. Read 3M+ as SKILL.md files collected from GitHub, not as distinct, maintained skills.

For programmatic use there is a free REST API and an MCP server. Browsing is free. Anonymous API access is capped at 50 requests per UTC day, free account keys get 100 keyword searches per day, and the Standard and Professional tiers allow 5,000 and 25,000 searches per month.

SkillsMP was also one of the two sources for a January 2026 security study of 31,132 skills (more on that below). Use it to find candidates, then judge them at the source.

skills.sh: Vercel’s Registry and the npx skills CLI

skills.sh is Vercel’s open directory for agent skills, launched on January 20, 2026 together with the skills CLI. If you searched for “skills sh”, this is it: the website is the index and leaderboard, the CLI is the installer, and the skills themselves stay in their GitHub repositories.

# install skills from a repo (interactive picker), or one skill by name
npx skills add vercel-labs/agent-skills
npx skills add vercel-labs/agent-skills --skill web-design-guidelines

# search, check for updates, update, remove
npx skills find react
npx skills check
npx skills update
npx skills remove web-design-guidelines

The CLI installs into the current project by default. Add -g to make a skill global and -a claude-code (or another agent name) to target one agent. It supported 17 agents at launch; the README now lists OpenCode, Claude Code, Codex, Cursor “and 75 more.”

Size and ranking. The skills.sh API, opened on June 5, 2026, covers more than 600,000 skills. The leaderboard ranks them by installs reported through anonymous CLI telemetry, according to the skills.sh docs. As of this update, Vercel’s own find-skills leads with 4.0M installs, Matt Pocock’s grill-me follows at 1.4M, and Anthropic’s frontend-design sits at 1.0M.

Security. Since February 17, 2026, skills get automated audits from Gen (Agent Trust Hub), Socket and Snyk. Skills flagged as malicious are hidden from search and the leaderboard, a direct link shows a warning, and from skills@1.4.0 the CLI shows audit results and a risk level before installing. The audits page puts the three verdicts side by side, and they often disagree: design-mobile-apps, a top-25 skill, is “Safe” on Gen and “High Risk” on Snyk.

One gotcha: setting DISABLE_TELEMETRY=1 or DO_NOT_TRACK=1 also turns off the CLI’s security-audit requests. Privacy-minded users who set those variables install without seeing the audit.

Anthropic’s Claude Skills Marketplaces (Claude Code Plugins)

There is no standalone “Claude skills marketplace.” Anthropic ships skills inside plugins, and a plugin marketplace is a git repository with a .claude-plugin/marketplace.json catalog. Claude Code registers Anthropic’s official marketplace, claude-plugins-official, the first time you start an interactive session, so this works out of the box:

/plugin
/plugin install commit-commands@claude-plugins-official

The first command opens the Discover tab, where you can search everything in your marketplaces. Most of the official catalog comes from partners and other authors, not Anthropic. The web view, Claude Marketplace, lists 342 plugins and marks some as “Anthropic verified,” including Frontend Design, Code Review and Feature Dev. Superpowers and Context7 sit near the top of the page. Before you choose a scope (user, project or local), the details pane lists every skill, hook and MCP server the plugin will add, and for official-marketplace plugins it estimates the context cost per turn.

Anthropic runs four more catalogs worth knowing:

• anthropics/skills (183k GitHub stars): the reference skills, including the docx, pdf, pptx and xlsx skills behind Claude’s file creation. Add it with /plugin marketplace add anthropics/skills, then run /plugin install document-skills@anthropic-agent-skills. The document skills are source-available rather than open source, and the repo labels everything as demonstration material.
• Community marketplace (claude-community): third-party plugins that their authors submitted to Anthropic. Add it with /plugin marketplace add anthropics/claude-plugins-community. Nearly every entry is pinned to a commit SHA, and Claude Code refuses to install a different commit.
• Anthropic’s directory at claude.ai/directory: the catalog for claude.ai and Cowork. Anyone on a paid plan can submit a plugin, and every version gets automated validation and a security scan, with a person reviewing each new listing. Plugins added there sync into Claude Code as <name>@synced.
• Demo marketplace (anthropics/claude-code): example plugins, most of them duplicated in the official marketplace. Older tutorials still point here; install from claude-plugins-official instead.

On claude.ai itself, you can upload a skill as a ZIP under Customize > Skills, and Team and Enterprise owners can provision skills for the whole organization.

The trust model deserves a careful read. Claude Code reserves Anthropic’s marketplace names for repositories under github.com/anthropics/, so a third-party catalog cannot pose as an official one. But plugin hooks, monitors and MCP servers run outside Claude Code’s sandbox with your user permissions, the install pane warns that Anthropic “does not control what MCP servers, files, or other software are included in plugins,” and Anthropic does not review third-party marketplaces at all. claude-plugins-official auto-updates by default; the community and third-party marketplaces don’t, which is the right call for code you reviewed once.

OpenAI: One Plugin Directory for Codex and ChatGPT

OpenAI took the same route as Anthropic: skills for local use, plugins for distribution. Codex reads skills from .agents/skills in every directory from your working directory up to the repo root, from $HOME/.agents/skills, and from /etc/codex/skills for admins. The built-in $skill-installer pulls curated skills by name ($skill-installer linear) and can fetch from other repositories when you ask it to.

The openai/skills catalog behind that installer is now marked deprecated, with a pointer to openai/plugins. The Codex skills docs are explicit: “For reusable distribution of your own skills, prefer plugins.” And per the plugin docs, “ChatGPT and Codex use the same public plugin catalog,” with tabs for OpenAI-built plugins, your workspace’s plugins and your own. Run /plugins in Codex CLI, or open the Plugins tab in the Codex or ChatGPT app. Workspace admins can import and sync a GitHub marketplace for their team.

Security posture: plugin capabilities run under Codex’s sandbox and approval policy, and the docs tell you to review plugin hooks before they run. OpenAI publishes neither a catalog size nor a review process for third-party listings. Skills trigger implicitly when a task matches their description or explicitly via $; set allow_implicit_invocation: false to make one manual-only. If you build on the API rather than in Codex, our Responses API skills post covers server-side skills.

GitHub gh skill: Install From Any Repo, With Pinning

GitHub added gh skill to its CLI on April 16, 2026, as a public preview that needs gh v2.90.0 or later. It has no catalog of its own: any GitHub repository with skills is the marketplace. What it adds is supply-chain hygiene the other installers lack.

gh skill search terraform
gh skill preview github/awesome-copilot documentation-writer
gh skill install github/awesome-copilot git-commit --agent claude-code --scope user
gh skill install github/awesome-copilot git-commit --pin v2.0.0

Installs write the source repository, ref and tree SHA into the skill’s frontmatter. Pinned skills are skipped by gh skill update, so upgrades happen when you decide. gh skill publish validates against the agentskills.io spec and offers to enable immutable releases. The man page lists more than 40 agent targets, from Copilot, Claude Code, Cursor, Codex and Gemini CLI to Goose, Junie and OpenClaw. GitHub’s own warning still applies: skills “are not verified by GitHub and may contain prompt injections, hidden instructions, or malicious scripts.”

Other Directories: ClawHub, Curated Lists and Paid Stores

• ClawHub is OpenClaw’s registry, run by the OpenClaw Foundation. Since February 7, 2026, every upload is hashed and checked with VirusTotal, including its Gemini-based Code Insight: malicious verdicts are blocked, suspicious ones get a warning, and active skills are rescanned daily. OpenClaw itself calls this “not a silver bullet,” and the registry’s history explains why: see why skill marketplaces keep getting poisoned.
• VoltAgent/awesome-agent-skills (36k stars) lists 1,000+ hand-picked skills from official teams like Anthropic, Stripe, Cloudflare, Supabase and Trail of Bits, plus the community. Curated means someone chose each entry, not that anyone audited it.
• Agensi is the paid model: it claims 13,000+ skills, takes a flat 30% platform fee (creators keep 70%), says each skill passes an “8-point security scan,” and delivers skills through an MCP server instead of files on disk.
• LobeHub and agentskill.sh run further aggregators with their own install tooling, per KDnuggets’ April roundup.

Vendor Skills: Install From the Source

Some of the most useful skills come straight from platform vendors. dbt Labs and Supabase both ship two install paths: the GitHub repo doubles as a Claude Code marketplace and as a npx skills package.

dbt Labs published official agent skills on February 5, 2026, covering analytics engineering (models, tests, source exploration), the semantic layer (metrics and semantic models with MetricFlow), platform operations (job failures, the dbt MCP server) and migration from dbt Core to the Fusion engine.

# Claude Code
/plugin marketplace add dbt-labs/dbt-agent-skills
/plugin install dbt@dbt-agent-marketplace

# other agents
npx skills add dbt-labs/dbt-agent-skills --global

Supabase published Postgres best-practice skills on January 21, 2026: rules in eight categories, prioritized from critical (query performance, connection management, security and RLS) down to low (advanced Postgres features), written for machine parsing. Install with npx skills add supabase/agent-skills, or in Claude Code with /plugin marketplace add supabase/agent-skills followed by /plugin install postgres-best-practices@supabase-agent-skills.

Install vendor skills from the vendor’s repository, not from a copy an aggregator turned up. Copies drift, and a fork with one extra line in SKILL.md looks identical in a search result.

Are Agent Skills Safe to Install?

Safer than in February, still not safe by default. Four findings from 2026 research shape how much to trust any marketplace’s badge:

• Vulnerabilities are common. Agent Skills in the Wild analyzed 31,132 skills collected from SkillsMP and skills.rest in December 2025. 26.1% had at least one vulnerability, 5.2% showed high-severity patterns suggesting malicious intent, and skills bundling executable scripts were 2.12 times more likely to be vulnerable than instruction-only skills.
• Scanners disagree. A repository-aware study of 238,180 skills found that on skills.sh, Gen’s Agent Trust Hub flagged 13.76% of the skills it checked, Snyk 7.69% and Socket 3.79%. Of 27,111 skills.sh skills checked by five scanners, only 33 were flagged by all five. When the authors checked whether flagged skills fit their surrounding GitHub project, only 0.52% of a 2,887-skill sample stayed suspicious. Audit badges are noisy in both directions.
• Dead links are an attack surface. The same study found 121 skills, 77 of them on skills.sh, pointing to abandoned GitHub repositories whose names could be re-registered by anyone.
• Scanning can be evaded. Unit 42 found five malicious ClawHub skills that stayed available between February and May 2026. One put its payload at the top of a README, followed by 22 MB of padding to push the file past scanner size limits.

What to do about it:

1. Read the SKILL.md and every file in scripts/ before installing. Instruction-only skills carry less risk.
2. Pin versions. gh skill install ... --pin, the commit-pinned claude-community marketplace, or a vendored copy committed to your repo all stop a reviewed skill from changing under you.
3. Install at project scope. Both npx skills and gh skill default to it; reach for global only for skills you’d trust everywhere.
4. Treat plugins as code. In Claude Code, plugin hooks and MCP servers run with your user permissions outside the sandbox, whichever marketplace they came from.
5. Leave auto-update off for third-party marketplaces in Claude Code (the default), and don’t disable the skills CLI’s telemetry unless you check audits another way.

For the wider supply-chain picture, including the ClawHavoc campaign, see AI agent skills are the new npm packages.

AGENTS.md vs. SKILL.md: Installed Isn’t the Same as Used

SKILL.md defines a reusable capability that travels across projects and agents. AGENTS.md is project context in your repository root: architecture, conventions, workflows.

The catch is that agents don’t always reach for a skill. In Vercel’s Next.js 16 evals (January 2026), a skill with default behavior scored 53%, the same as no documentation at all, because the agent never invoked it in 56% of cases. Explicit instructions to use the skill raised the score to 79%. An 8KB docs index embedded in AGENTS.md hit 100%.

Use both. AGENTS.md for knowledge the agent needs on every turn of this project. Skills for task-specific workflows that should load only when relevant, and that you can also trigger by name.

Publishing Your Own Skill

Where you publish decides who can install it:

• skills.sh and SkillsMP: put the skill in a public GitHub repo. skills.sh ranks skills by installs reported from the CLI, so the route onto its leaderboard is people running npx skills add your-org/your-repo; npx skills init scaffolds a new skill. SkillsMP’s crawler guarantees neither inclusion nor a fixed update schedule.
• Claude: skills are not a submission type on their own. Bundle them in a plugin and submit it at claude.ai/directory/manage from a paid plan, or publish your own marketplace.json in a repo (publishing guide).
• Codex and ChatGPT: package the skill as a plugin, then share it through a marketplace source such as a repo marketplace for your team.
• GitHub: gh skill publish --dry-run validates your skills against the spec before you publish.

Frequently Asked Questions

What is SkillsMP?

SkillsMP is an independent search index for Agent Skills. It crawls public GitHub repositories for SKILL.md files and claims more than 3 million, searchable by keyword, category and occupation. It is not affiliated with Anthropic or OpenAI, does not install skills and does not security-review them. You install from the skill’s source repository after reading it.

What is skills.sh?

skills.sh is Vercel’s open directory for Agent Skills, paired with the skills CLI (npx skills add owner/repo). Launched in January 2026, it indexes more than 600,000 skills, ranks them by installs, installs into nearly 80 agents including Claude Code, Codex and Cursor, and shows security audits from Gen, Socket and Snyk before installation.

Is there an official Claude skills marketplace?

Yes, in the form of plugin marketplaces. Claude Code registers Anthropic’s official marketplace, claude-plugins-official, automatically; browse it with /plugin or on Claude Marketplace, which lists 342 plugins. Anthropic’s reference skills install with /plugin marketplace add anthropics/skills, and claude.ai has its own directory, where every submitted plugin version gets a security scan and new listings get human review.

How do I install agent skills in Claude Code?

Three ways. From a plugin marketplace, run /plugin install name@marketplace (add non-official marketplaces first with /plugin marketplace add owner/repo). With a cross-agent installer, run npx skills add owner/repo -a claude-code or gh skill install owner/repo skill-name --agent claude-code. Or copy the skill folder into .claude/skills/ in your project, or ~/.claude/skills/ for all your projects.

Are agent skills safe to install?

Not by default. A study of 31,132 skills from SkillsMP and skills.rest found 26.1% had at least one vulnerability. skills.sh, ClawHub and Anthropic’s directory now scan skills, but scanners disagree and can be evaded, and SkillsMP and gh skill do not scan at all. Read the SKILL.md and any scripts, pin versions, install at project scope, and treat plugins with hooks or MCP servers as code that runs with your permissions. For enterprise rollouts, see the supply chain security implications.