// documentation

skillgraph docs

The package manager for AI-agent skills — npm, but for skills. Pull a skill and its entire dependency graph into .claude/skills/ with one command, and publish your own under your @scope.

Overview

A skill is a flat-named folder with a SKILL.md (its name and description in frontmatter), following the open Agent Skills standard. Skills can declare dependencies on other skills, so the registry forms a directed graph — its namesake. Install a skill and skillgraph resolves the whole closure, pins it in a lockfile, and writes every skill into your project's .claude/skills/.

Concepts

Coordinates & scope: A published skill is addressed by a coordinate @scope/name. The scope is your username; it is stamped server-side from your identity, never typed — so you can only ever publish under your own. On disk the scope disappears: a skill always materializes to a flat .claude/skills/<name>/.
Owned vs installed: Installed skills are ones you consume (read-only). Owned skills are ones you manage and can publish. fork turns an installed skill into an owned copy under your scope, recording its origin as lineage.
The registry is the source of truth: Ownership, lineage, and content are server-authoritative. Your local files — the manifest, the lock, the per-skill sidecar — are regenerable, non-authoritative state; editing them can't forge ownership or inject content. Published versions are immutable.
Manifest → lock → skills: skillgraph.json is intent (the skills you want, by version or latest); skillgraph.lock.json is the resolved, pinned graph you got; .claude/skills/ is the materialized content. A bare install reproduces from the lock without re-resolving — npm's install vs ci, over a flattened, single-version graph.

CLI reference

Search and install need no account; sign in only to publish. The surface is scriptable and non-interactive — built for humans at a terminal and agents in a loop.

Consume

skg search <keyword> find skills by name or description
skg install @scope/name resolve the full closure into .claude/skills/
skg install reproduce the project from its lockfile
skg ls print the installed dependency tree
skg ls @scope/name inspect a coordinate before installing

Publish

skg init <dir> bring a SKILL.md folder under management
skg publish <target…> publish a graph of managed skills, leaves first
skg fork @scope/name scaffold an owned copy of another author's skill
skg revert @scope/name flip a fork back to its origin
skg import <git-url> clone a repo and publish its skills under your scope

Account

skg login sign in with GitHub, store an API key
skg whoami show the @scope you publish under
skg logout remove the stored key

FAQ

Do I need an account to install skills?: No. Search, install, and inspect are fully anonymous — you only sign in to publish your own skills.
Where do skills get installed?: Into your project's .claude/skills/ folder — Claude Code's skill-discovery directory — each as a flat <name>/ folder. The manifest (skillgraph.json) and lockfile (skillgraph.lock.json) live at the project root.
What is a coordinate and a scope?: A published skill is addressed as @scope/name, where the scope is your username. You never type your scope when publishing — the registry stamps it from your identity, so you can only ever publish under your own.
Owned vs installed — what is the difference?: Installed skills are ones you consume (read-only). Owned skills are ones you manage and can publish. fork turns an installed skill into an owned copy under your scope, recording its origin as lineage.
How does publishing a dependency graph work?: publish resolves the managed skills you point it at, orders them leaves-first, and publishes each. A co-published sibling dependency is stamped with your scope; a cross-author dependency is referenced by its full coordinate.
Is it Claude-only?: The on-disk target is Claude Code's .claude/skills/ for now, but the manifest and lock live at the project root (agent-agnostic) so the same resolved graph can materialize into other agents' directories later.