From Anthropic's SKILL.md, to Google's official google/skills, to Vercel's npx skills with skills.sh, skills now form a cross-vendor open package-management ecosystem. In one line: this is "npm for skills" — it solves discovery and getting a skill's files into each agent's own directory. (Table facts are public sources, cited below.)
| Tool | Positioning (public sources) | Through an asset mindset |
|---|---|---|
| npx skills (Vercel Labs) | Open pack manager: add / list / find / update, installing into .claude/.cursor/.github etc. | Install and done; no personal inventory, no deduplication |
| Anthropic / Google official CLIs | Publish and distribute official SKILL.md; Google ships product usage as skills | One-way distribution; never answers "what have you actually accumulated" |
| OpenSkills & registries | Cross-platform loading and discovery (skills.sh, agentskill.sh, etc.) | Adaptation and search; no tagging or dedup of your own assets |
| Flint (this project) | Manages all your skills from an asset perspective | Single local source of truth, tagging, search, dedup, on-demand deployment |
A quick look at popular skill tooling (public sources)
- npx skills (Vercel Labs)Open agent-skills package manager. `npx skills add <owner/repo>` installs a skill into the agent-specific directory (.claude/skills, .cursor/skills, .github/skills, etc.) via symlink or copy, with list / find / update / init; the skills.sh registry it ships with catalogs tens of thousands of skills.
- The standard setters and official CLIsAnthropic's official skills CLI (@anthropic-ai/skills) installs into .claude/skills; Google then released its official Agent Skills repo google/skills, where one line `npx skills add google/skills` distributes product knowledge as SKILL.md to any agent — turning SKILL.md into an interoperability standard.
- Community tools and registriesOpenSkills generates an AGENTS.md to load skills on non-Claude agents; catalog sites like skills.sh (Vercel, 91,000+ indexed), agentskill.sh (100K+), and officialskills.sh (581 curated) handle discovery and distribution.
The common thread is clear: these tools focus on fetch-and-install — pulling someone else's skill onto your machine, into a location the agent you're using can read. None of them own the harder question of "what do I actually have, which of them overlap, and which are my own". Because every install lands in an agent's own directory, skills stay scattered, and once you have enough of them you can no longer say what capabilities you've actually accumulated.
Flint differs on the fundamental view: it treats a skill not as a "software package to install" but as a personal asset that belongs to you. Its emphasis moves from downloading to accumulating and using — a single local source of truth, with tagging, search, and deduplication, so you end up with your own inventory of capabilities. Deployment is just a mount of that asset; it never changes ownership and never creates a second copy.
An install command solves discovery, not personal knowledge assets. The real question isn't "what can I install" but "what do I actually have, and who does it belong to". Tools answer the former; an asset mindset answers the latter.
In one sentence: the npx skills family is "npm for skills" — it answers "what can I install"; Flint is closer to "Obsidian for skills" — it answers "what do I actually have, and it's mine". They aren't mutually exclusive: use npx skills to pull a new skill in, then hand it to Flint to deduplicate and fold into your asset inventory.
Sources · 参考来源