Agent Skill Marketplaces: Where to Find and Share Skills in 2026
The agent skill ecosystem crossed 400,000 published skills in early 2026. That number keeps climbing. Every week, developers publish new SKILL.md files for Claude Code, Codex, and OpenCode — covering everything from database migrations to infrastructure provisioning to content generation.
The problem is no longer supply. It is discovery. Finding the right agent skill marketplace, evaluating what you find there, and knowing whether an installed skill actually works in your environment — these are the real challenges facing developers today.
This guide compares the six major places to find, publish, and install agent skills in 2026. We cover what each marketplace offers, where each falls short, and what quality signals you should look for before installing anything.
What Is an Agent Skill Marketplace
An agent skill marketplace is any platform where developers discover, share, and install skills for AI coding agents. Skills are structured instructions — typically SKILL.md files — that extend an agent's capabilities for specific tasks. A skill might teach an agent how to generate Terraform configs, write API documentation, or run security audits.
Marketplaces vary in scope. Some are dedicated registries built specifically for skill distribution. Others are general-purpose platforms (GitHub, npm) where skills coexist alongside millions of other packages and repositories. A few are platform-locked, tied to a specific agent like Claude Code or Codex.
The differences matter. A dedicated registry can enforce quality standards, run security audits, and surface trust signals. A general-purpose platform offers reach and familiarity but no curation. Understanding these tradeoffs helps you choose where to search — and where to publish.
Top Agent Skill Marketplaces Compared
Here is the current landscape. Each marketplace serves a different segment of the ecosystem.
GitHub remains the de facto agent skill marketplace by volume. Searching for SKILL.md files returns hundreds of thousands of results. Many popular skills live in standalone repositories or inside larger project repos. GitHub provides stars, forks, and commit history as proxy quality signals — but no skill-specific curation or verification. You clone or copy what you find. Quality varies wildly.
skills.sh is the first dedicated skill registry with built-in audit infrastructure. Skills submitted to skills.sh undergo automated security analysis and receive trust badges when they pass. Installation uses a familiar CLI pattern: npx skills add author/skill-name. The catalog is smaller than GitHub but growing fast, and every listed skill meets baseline quality thresholds that GitHub does not enforce.
npm and npx serve as a distribution layer for skills packaged as Node modules. Some skill authors publish their SKILL.md files alongside tooling as npm packages. The npm ecosystem brings download counts, versioning, and dependency management — infrastructure that skill-only registries are still building. The tradeoff: npm was not designed for skills, so discoverability requires knowing what to search for.
Claude Code skill-creator is Anthropic's built-in skill authoring and distribution system. Skills authored through skill-creator can be shared within the Anthropic ecosystem. Anthropic reviews submitted skills and provides eval scores based on their internal benchmarks. The limitation: skills published here work only with Claude Code. If you also use Codex or OpenCode, you need skills from a platform-agnostic source.
OpenClaw is an open-source agent platform with a community skill-sharing layer. Skills on OpenClaw receive community ratings and usage statistics. The catalog skews toward infrastructure and DevOps tasks, reflecting the platform's user base. Installation uses the OpenClaw CLI: opncl install author/skill. The community is smaller but active, with strong norms around documentation quality.
Codex skill registry is OpenAI's distribution channel for Codex-compatible skills. Like Anthropic's skill-creator, it offers platform-vendor review and evaluation scores. Skills here are optimized for Codex's execution model and may not transfer cleanly to other agents. The catalog has grown steadily since Codex added skill support in late 2025.
Agent Skill Marketplace Comparison Table
This table summarizes the key differences across all six marketplaces:
| Marketplace | Platform Support | Curation | Install Method | Skill Count | Quality Signals |
|---|---|---|---|---|---|
| GitHub | All agents | None | Manual clone | 200K+ | Stars, forks, commit recency |
| skills.sh | All agents | Security audited | npx skills add |
50K+ | Audit badge, maintainer verified |
| npm | All agents | None | npm install |
100K+ | Download count, version history |
| Claude skill-creator | Claude Code only | Anthropic reviewed | Built-in | 30K+ | Anthropic eval scores |
| OpenClaw | OpenClaw only | Community moderated | opncl install |
10K+ | Community ratings, usage stats |
| Codex registry | Codex only | OpenAI reviewed | Built-in | 20K+ | OpenAI eval scores |
Skill counts are approximate estimates based on ecosystem size and growth rates as of March 2026.
How to Choose Skills from a Marketplace
Finding a skill is the easy part. Evaluating whether it will work for your use case takes more effort. Here is what to check before installing.
Read the SKILL.md thoroughly. A well-written SKILL.md includes a clear description of what the skill does, when it should trigger, what inputs it expects, and what outputs it produces. If the SKILL.md is vague or missing, move on. Poorly documented skills are poorly maintained skills.
Check quality signals relevant to the marketplace. On GitHub, look at stars and recent commit dates. A skill with 2,000 stars but no commits in six months may not work with current model versions. On skills.sh, look for the security audit badge. On npm, check weekly downloads — sustained download counts indicate active use, not just initial curiosity.
Verify platform compatibility. A skill built for Claude Code may reference tool-use patterns that Codex does not support. Check whether the SKILL.md uses platform-specific features or follows the platform-agnostic SKILL.md standard. Cross-platform skills are more valuable and tend to receive more community testing.
Look for tests or evals. Some skill authors include eval sets — test prompts and expected behaviors. Skills with evals demonstrate that the author thought carefully about edge cases. Skills without evals require you to do that thinking yourself.
Check the skill's trigger coverage. The description field in SKILL.md determines when the agent activates the skill. A description that only matches exact phrases ("generate a PowerPoint") will miss natural language variations ("make me some slides for the board meeting"). Broad, well-crafted descriptions produce better trigger rates.
Publishing Your Skills to a Marketplace
If you have built skills worth sharing, publishing them increases their reach and invites community feedback.
Structure your SKILL.md for discoverability. The description field is the single most important line for marketplace search. Write it to cover the range of queries a user might send when they need your skill. Include both the technical term and the natural language equivalent. "Generates Terraform infrastructure-as-code configurations from architecture descriptions" covers more search surface than "Terraform skill."
Follow the SKILL.md standard. Every marketplace indexes the same core fields: name, description, triggers, and body. Deviating from the standard format means some marketplaces may parse your skill incorrectly or exclude it from search results entirely.
Cross-post across multiple marketplaces. Publishing to GitHub, skills.sh, and npm simultaneously maximizes your audience. The SKILL.md format is platform-agnostic by design — the same file works everywhere. Use GitHub as your source of truth, then publish to skills.sh for the audit badge and npm for the package ecosystem.
Include an eval set. Even a small eval set of 10-20 test prompts significantly increases trust. It shows prospective users that you have tested the skill against real scenarios and care about quality beyond the initial publish.
The Missing Layer: Skill Quality Infrastructure
Marketplaces solve discovery. They tell you what exists. They do not tell you what works.
No marketplace currently reports whether a skill triggers reliably in production. No marketplace tracks pass rates, false positive rates, or performance regressions across model updates. Stars and download counts measure popularity, not effectiveness.
This is the gap between finding a skill and trusting a skill. A marketplace can show you that a Terraform skill exists, has 5,000 stars, and was updated last week. It cannot show you that the skill triggers on only 64% of relevant queries in your environment, or that the latest model update caused a 15% regression in output quality.
selftune adds the quality layer that marketplaces are missing. It monitors whether installed skills actually trigger and execute correctly against real session data. It detects missed triggers, proposes improved descriptions, validates changes against eval sets, and rolls back automatically if quality drops. Marketplaces get skills onto your machine. selftune tells you whether they are working once they are there.
npx selftune@latest doctor
# Diagnoses skill health: trigger rates, missed invocations, description coverage
Think of it this way: marketplaces are the app store. selftune is the performance monitoring you install after downloading the app.
Frequently Asked Questions
Which marketplace has the most agent skills?
GitHub has the largest catalog by volume, with over 200,000 SKILL.md files indexed across public repositories. However, this includes unmaintained, duplicate, and low-quality entries. For curated selections, skills.sh and the platform-specific registries (Claude skill-creator, Codex registry) offer smaller but higher-quality catalogs.
Can I use skills from multiple marketplaces at once?
Yes. The SKILL.md format is platform-agnostic. You can install a skill from GitHub, another from skills.sh, and a third from npm — all running in the same agent environment. The key constraint is agent compatibility, not marketplace origin. A skill built for Claude Code works in Claude Code regardless of where you downloaded it.
How do I know if a marketplace skill is safe to install?
Start with marketplace-level signals. On skills.sh, look for the security audit badge — these skills have passed automated analysis for common risks like credential exposure and file system abuse. On GitHub, check the repository owner, contributor history, and open issues. On npm, look for established publishers with multiple packages. For any marketplace, read the SKILL.md body before installing. A skill that requests broad file system access or network permissions deserves extra scrutiny.
How do I know if an installed skill actually works well?
Without observability tooling, you do not. Marketplaces track installs, not performance. You can check manually by testing a skill against sample prompts, but this only covers cases you think to test. For continuous monitoring, tools like selftune track trigger rates, detect missed invocations, and measure execution quality against real session data — giving you the production visibility that marketplaces do not provide.