SkillAtlas — أطلس المهارات

واجهة عربية/إنجليزية سريعة لاكتشاف مهارات وكلاء الذكاء الاصطناعي، مبنية على بيانات ComposioHQ/awesome-claude-skills.

SkillAtlas turns the upstream Markdown catalog and bundled SKILL.md files into a searchable, production-ready directory without executing upstream code.

What is included

• 997 indexed records in the current snapshot: 864 real SKILL.md files plus reconciled README entries.
• Instant Arabic/English search with category, source, type, script, and favorite filters.
• 997 statically generated skill pages with source, licensing, resource flags, and manual install guidance.
• Arabic-first RTL UI, English toggle, dark/light modes, responsive navigation, loading/error/empty states.
• Atlas Zero agent with deterministic catalog reasoning, safe in-site actions, and no model API key.
• Optional public research across GitHub, npm, and Arabic/English Wikipedia with visible sources.
• Optional Qwen3-0.6B local brain in a Web Worker using WebGPU; it downloads only after explicit consent.
• Agent Elements chat interface from 21st.dev with local browser conversation history.
• SEO metadata, JSON-LD, Open Graph image, sitemap, robots, web manifest, and catalog status API.
• Deterministic parser tests and a twice-daily GitHub Actions sync that commits only after tests and build pass.
• Production dependency overrides for patched postcss and sharp versions.

Run locally

Requirements: Node.js 20.9 or newer and npm.

npm install
npm run dev

Open http://localhost:3000.

Production verification:

npm run check
npm run build
npm start

Refresh the catalog

Clone or update the upstream repository, then point the sync script to that checkout:

npm run sync -- --source ../awesome-claude-skills

Windows PowerShell example:

npm.cmd run sync -- --source ..\awesome-claude-skills

The generator writes:

• src/data/skills.generated.json — normalized, search-ready records.
• src/data/snapshot.generated.json — source commit, counts, categories, and reconciliation warnings.

Generation is deterministic. If the upstream commit has not changed, the existing syncedAt value is preserved and no data diff is created.

Automatic update workflow

.github/workflows/sync-upstream.yml runs twice every day and can also be started manually. It:

1. clones a shallow upstream snapshot;
2. parses every real SKILL.md and the README catalog as separate sources;
3. reconciles exact paths and keeps unresolved README listings without inventing broken file links;
4. runs parser tests, ESLint, TypeScript, and the full Next.js build;
5. commits generated JSON only when all checks pass and content changed.

No script, package, or executable from the upstream repository is run during synchronization.

Architecture

Upstream README + 864 SKILL.md files
                │
                ▼
      scripts/sync-skills.mjs
                │
        deterministic snapshot
                │
      ┌─────────┴─────────┐
      ▼                   ▼
Server Components    Client directory
SSG detail pages     search + filters

Main routes:

• / — landing and curated discovery.
• /explore — full searchable directory.
• /skills/[slug] — static skill details.
• /learn — Skills vs Tools vs MCP and safe install guidance.
• /about — methodology, attribution, limitations, and current snapshot.
• /agent — keyless agent, optional live research, and optional local WebGPU brain.
• /privacy — clear data-flow, storage, research, and safe-use disclosure.
• /api/catalog/status — generated catalog metadata.
• /api/agent — bounded deterministic agent endpoint.
• /api/research — allowlisted public research endpoint with request limits and no response cache.

Deploy

The project is ready for Vercel or any Node.js host that can run the standard Next.js production server.

Set this variable to the production origin before building:

NEXT_PUBLIC_SITE_URL=https://your-domain.example

Vercel needs no additional build configuration: install with npm install, build with npm run build.

Atlas Zero needs no paid AI credential. The optional local model is downloaded from Hugging Face by the visitor's browser after consent (roughly 570 MB plus tokenizer/support files), then cached and executed locally. Public research queries are sent only to the providers named in the interface.

Source, safety, and licensing

SkillAtlas is an independent interface and is not an official Anthropic or Composio product.

The upstream README states Apache-2.0 for the repository while warning that individual skills may have different licenses. The snapshot therefore keeps per-skill license metadata and does not assume that every file can be redistributed. Full skill text is rendered in SkillAtlas only when a recognizable open license is detected; otherwise the detail page links to the original source.

Presence in this directory is not a security audit or compatibility guarantee. Review SKILL.md, scripts, dependencies, requested tools, and permissions before installing a skill.