Built on the open Agent Skills standard

The package manager
for agent skills.

Manage groups of related skills like the dependencies they are. Resolved, versioned, and installed flat into .claude/skills/ with one command.

$ npm i -g skillgraph

// why skillgraph

A package manager that understands
how skills depend on skills.

Everything you expect from a modern registry — resolution, versioning, search, reproducibility — purpose-built for the way agents consume skills.

Dependency resolution on an open standard

Every skill is a plain SKILL.md folder per the agentskills.io standard — skillgraph adds the dependency graph on top. Skills depend on skills; it resolves the full transitive closure, detects cycles and dangling edges, and installs leaves-first, the npm model for agents, without ever rewriting your content.

Scoped, addressable, offline-first

Skills are addressed by @scope/name coordinates and versions never change. Author and wire them together with no registry — your local skills carry an @local scope the server stamps to @you when you publish. Your scope comes from your identity server-side, so you can only ever publish as yourself.

Search by name or description

The registry indexes the description from each skill’s frontmatter, so you find skills by what they do — not just what they’re called.

// the CLI

One tool, the whole lifecycle

Consume skills and publish your own from one scriptable, non-interactive surface that runs clean for humans and agents alike.

Consume skills
  • skg install @scope/name resolve the closure into .claude/skills/
  • skg ls print the installed dependency tree
  • skg ls @scope/name inspect a coordinate before installing
  • skg search <kw> find skills by name or description
Publish & fork skills
  • skg init <dir> bring a SKILL.md under management
  • skg publish <dir> publish a graph of managed skills
  • skg fork @scope/name fork another author’s skill under your @scope

// get started

Install the CLI and
pull your first skill.

Install, and you’re installing skills in seconds — no setup. Log in to publish your own. Works for humans at a terminal and agents in a loop.