About AI-SKILLS

The open ecosystem for distributing agent skills and plugins.

The problem

The agent ecosystem is fragmented. Skills live in one registry, plugins in another, and installing either requires jumping between CLIs that don't talk to each other. None of them truly fit how developers work.

We ran into this ourselves. Running --help on a subcommand shouldn't execute a destructive action — but it did, silently overwriting local modifications. Removing a skill for one agent left ghost copies in shared directories that other agents still read. SSH and GitLab sources broke on updates because the CLI naively appended folder paths to clone URLs. Plugin installation required a separate authenticated session and a proprietary backend that couldn't run offline.

These aren't edge cases — they're daily friction for anyone managing skills across a team or more than one coding agent.

The landscape

skills.sh / Vercel npx skills — the most popular skills-only CLI. Git-native for installation, but no plugin support, no unified manifest, and relies on a skills-lock.json that only tracks skills (not plugins or team configs). Discovery and metadata are coupled to the skills.sh backend.

skills.sh Platform CLI ai — extends skills.sh with plugin support, but requires login, a proprietary registry API for all non-git operations, and only installs plugins to Claude Code and Cursor. Can't work offline or against private infrastructure without their backend running.

aipack — a profile-based composition engine with a sync model. Different paradigm: you author “packs” with a pack.json manifest and sync them into harness-native configs. Powerful for complex setups, but not git-native distribution — packs are local artifacts, not published repos.

agentpkg — a bundle-first approach that compiles a .agentpkg.zip into platform-specific directory structures. Good for single-agent portability, but the zip-and-compile model doesn't support team collaboration, version tracking, or incremental updates.

Each tool solves a real problem. None of them unify skills and plugins under a single, git-native, reproducible workflow that works across all agents without a proprietary backend.

Our approach

Git-native distribution — every skill and plugin is a git repo. No proprietary hosting, no registry to sign up for. Works with GitHub, GitLab, self-hosted — if you can git clone it, you can install from it.
ai-package.json — one manifest to declare skills, plugins, and team configs. Version pins are embedded directly (ref@commitSha) — no separate lock file to keep in sync.
One CLI, all agents — install to Cursor, Claude Code, Codex, Copilot, Windsurf, or 40+ other agents from a single ai-pkgs command.
Predictable CLI behavior — --help never executes side effects. --ai mode fails loudly with the exact flag you need, not silently. Per-agent directories mean no ghost copies after removal.
Workspace workflow — fork a skill into your project, iterate locally, push back upstream. Like npm link for agent skills.
Free forever — open-source CLI, open registry, zero fees. We will never charge for distributing skills.

Where we are today

AI-SKILLS is young. Plugin installation currently supports Claude Code, Cursor, and Codex. Marketplace search is still in development. Our leaderboard data builds on top of the skills.sh API — we acknowledge standing on their shoulders.

What we're committed to: the ai-package.json format stays compatible, the CLI stays free, and the distribution model stays git-native. Everything else is iterating fast.

Contact

@foxsnowing gsnowingfox@gmail.com