Your AI assistant shouldn't install random prompt files

What it does

Agent Skills packages reusable instructions—SKILL.md files, templates, and references—into a managed catalog you can extend across AI coding agents like Cursor, Claude Code, and Copilot. The project treats these prompt packs as supply-chain artifacts: every skill is scanned with Snyk Agent Scan before publishing, hashed for integrity, and distributed via a CLI that enforces path isolation, symlink guards, and an audit trail. A local cache keeps the catalog available offline.

The interesting bit

The project essentially applies npm-style package management paranoia to AI prompts. It even exposes the catalog through an MCP server with “progressive disclosure”—agents search the index first, then fetch only the specific skill files they need, rather than slurping down an entire marketplace of untrusted text.

Key highlights

• Claims to counter a Snyk report finding that over 13% of marketplace skills contain critical vulnerabilities, positioning itself as a hardened alternative.
• Supports a wide range of agents, from Tier 1 tools like Cursor and GitHub Copilot to enterprise options like Amazon Q and Sourcegraph Cody.
• Every skill is open source with no binaries, backed by static analysis in CI/CD, immutable lockfiles, and content hashing.
• Includes an MCP server that lets AI agents query the catalog directly without downloading the full registry upfront.
• Skills cover practical domains: spec-driven development, AWS architecture, Playwright automation, Figma-to-code translation, and security reviews.

Verdict

Worth a look if you’re feeding third-party prompts into AI agents and want some supply-chain hygiene. Probably overkill if you only use a single agent with a handful of homegrown instructions.