AgentSkills CLI vs vercel-labs/skills

AgentSkills CLI is a faithful .NET-native port of vercel-labs/skills that adds first-class NuGet and npm support plus a layer of ergonomics. This page is the full feature-by-feature table; the FAQ entry summarizes the highlights in prose.

Capability matrix

Capability	AgentSkills CLI	npx skills
NuGet packages as a first-class source (public + private feeds, NuGet.config + credential providers)	yes	no NuGet path
npm registry fetch as a first-class source (public + private, .npmrc + scoped registries + _authToken)	yes	only experimental_sync from pre-installed node_modules
Skills can ship inside existing library packages (drop skills/ into your .nupkg or .tgz - same package id, version-locked with the SDK, discoverable via standard package search)	yes	n/a - no NuGet path; npm only via experimental sync
Version-aware target matching - remove Pkg@1.5.0 strict-matches; a "version 1.4.0 is installed" hint when the pin is wrong	yes	no concept of pinned matching
--path works across every source (NuGet, npm, git, local) - overrides where discovery scans inside the staged source	yes	only /tree/<ref>/<path> in GitHub URLs
Unified positional targets on list / remove - each arg matches as a skill name or any source format, union semantics	yes	separate positionals/flags
--by package|path|agent|scope grouping on list	yes	flat output
Install paths surfaced everywhere - Path column on add output, Installed under ... summary, --paths on list	yes	success status without paths
Multi-targeted runtime - LTS .NET 8 and latest .NET 10 from one .nupkg	yes	Node 18+ (single ecosystem)
DI extension points - register ISkillSourceFactory for new source types (cargo, oci, conda, ...) and ISkillSearchProvider for new search backends (internal registries, GitHub topic search, ...) with no core changes	yes	procedural, no public extension contract
Typed exception hierarchy with carried ExitCode - clean single-line errors, no stack-trace spam	yes	ad-hoc throws
Verbosity control - -v / -q / --trace route through Microsoft.Extensions.Logging to a Spectre logger	yes	console writes only
0 build warnings, TreatWarningsAsErrors, NetAnalyzers + Meziantou.Analyzer at Recommended	yes	not enforced

Where the two tools are identical

By design - the two tools interoperate, so being identical here is a feature, not a gap:

• SKILL.md format - same YAML frontmatter schema (name, description, optional metadata, optional internal flag)
• Lock file shapes - both use ~/.agents/.skill-lock.json (v3) for global state and ./skills-lock.json (v1) for project state, so installs done by either tool are visible to the other
• Universal .agents/skills/ install directory - shared by every agent that follows the universal convention
• Well-known endpoint - both support /.well-known/agent-skills/index.json (modern) and /.well-known/skills/index.json (legacy fallback)
• GitHub auth chain - both fall back through GITHUB_TOKEN -> GH_TOKEN -> gh auth token after an unauthenticated 403

What AgentSkills CLI v1 deliberately doesn't ship

Feature	Why it's not in AgentSkills CLI
Telemetry	Privacy by default. Lock files give you everything you need to track usage internally.
experimental_install	Still experimental in npx skills; holding for stability.
experimental_sync	Same. AgentSkills CLI' npm fetch covers the actual use case (install from registry).
The full 55-agent registry	v1 ships the most-used 5: claude-code, codex, cursor, opencode, universal. Adding more is mechanical - open an issue for any specific agent you need.
Live fzf-style find UI	The current table-output find is responsive enough; a TUI variant may land in a later version if there's demand.

Coexistence

If you have both tools installed, they coexist by design - different binaries (skills vs agentskills-cli), shared lock and install dirs. See Troubleshooting -> coexistence for the details.

Next

• FAQ - common questions with concise answers
• Source formats - every input type AgentSkills CLI accepts
• Search providers (extension point) - the contract for custom search backends