The most complete, most efficient capability pack for Claude Code — from web design to cybersecurity. Ships expert-level skills for everything, yet adds near-zero startup token cost because Claude routes from a tiny map and loads exactly one file per task.
⭐ If this saves you tokens or time, star it — it genuinely helps others find it.
Big skill packs have a hidden cost: dump hundreds of skills into Claude and every session pays for hundreds of descriptions in context. Claude Arsenal solves both sides:
- 🧠 Expert on everything. Every skill and agent is framed to operate at principal / senior level — even a one-line change gets senior judgment, not a lazy shortcut.
- ⚡ Token-minimal by design. The only thing that auto-loads is a ~65-line routing map. 2,247 skills sit on disk and load one at a time, on demand. Opening a new terminal costs ~1 small file — not 2,247 descriptions.
- 🧷 Remembers itself. Drop it in a folder and Claude Code auto-loads its
CLAUDE.mdthe moment youcdin. No install step, no re-reading the tree every session. - 📦 Self-contained & shareable. Pure Markdown + a tiny Node generator. Zip it, fork it, drop it in any project. MIT-licensed.
flowchart LR
A[New terminal in folder] --> B[Auto-load CLAUDE.md<br/>~65 lines = the cache]
B --> C{Task matches<br/>domain map?}
C -- yes --> D[Read ONE skill file<br/>skills/name/SKILL.md]
C -- niche --> E[grep CATALOG.md once] --> D
C -- specialist --> F[Spawn ONE agent]
D --> G[Do the work]
F --> G
style B fill:#10b981,color:#fff
style D fill:#3b82f6,color:#fff
style G fill:#8b5cf6,color:#fff
The 2,017-file community library and the full INDEX.md are never bulk-read — Claude greps a
single catalog line, then opens one file. That's the whole trick.
| Count | Notes | |
|---|---|---|
| 🤖 Agents | 66 | Principal-level specialists, spawn on demand |
| 🎯 Core skills | 230 | Hand-authored, expert-framed, consistent (19 domains) |
| 🌍 Community skills | 2,017 | Imported from open source, deduplicated (6,443 → 2,017) |
| 📚 Total skills | 2,247 | Loaded one file per task |
| 🧭 Auto-load footprint | ~65 lines | The routing map, nothing else |
🎨 Web Design · ⚛️ Frontend · 🔧 Backend · 🗄️ Databases · 🚀 DevOps · ☁️ Cloud · 🛡️ Cybersecurity (25) · 🤖 AI/ML · 📱 Mobile · ✅ Testing · 🔀 Data Engineering · 🔌 API/Integration · ⚡ Performance · 📝 Docs · 📋 Project Mgmt · 🌿 Git · 💻 Languages · ⛓️ Blockchain · 🎮 Game Dev
# 1. Get it
git clone https://github.com/fkfkg/Agent-Library.git
cd Agent-Library
# 2. Open Claude Code in this folder — CLAUDE.md auto-loads. That's it.
# Or install into another project:
./scripts/install.sh /path/to/your/project # macOS/Linux
./scripts/install.ps1 -Target C:\path\to\project # Windows PowerShellThen just ask — Claude routes to the right expert automatically:
| You ask | Claude routes to |
|---|---|
| "Build a responsive pricing page" | responsive-layout + landing-page-design |
| "Audit this auth code" | security-auditor agent + authentication-security |
| "This query is slow" | sql-optimizer agent + query-optimization |
| "Set up a canary deploy" | canary-releases + devops-engineer |
claude-arsenal/
├── CLAUDE.md # ~65-line routing map (the only thing that auto-loads)
├── INDEX.md # full core catalog — opened on demand
├── manifest.json # machine-readable catalog
├── agents/ # 66 agent definitions
├── skills/ # 230 curated core skills (<name>/SKILL.md)
├── skills-community/ # 2,017 imported+deduped skills + CATALOG.md + ATTRIBUTION.md
└── scripts/ # generator, importer, validator
# Edit scripts/skills-data.mjs or scripts/agents-data.mjs, then:
node scripts/generate.mjs # regenerates core skills, agents, INDEX.md, manifest.jsonThe community library is imported separately (see scripts/IMPORT.md) and left
untouched by the generator.
skills-community/ aggregates open-source skills, deduplicated to the single best copy per topic.
Full source + license breakdown: skills-community/ATTRIBUTION.md
(every skill file also carries its origin in a footer). Sources — all MIT / Apache-2.0:
sickn33/agentic-awesome-skills, addyosmani/agent-skills, mattpocock/skills, wshobson/agents,
coreyhaines31/marketingskills, anthropics/skills. Anthropic's source-available document skills
were excluded from redistribution.
Offensive-security content (penetration-tester, red-team-operator, malware-analyst, pentest
skills) is for defensive and authorized testing only — your own systems, scoped engagements,
CTFs, or isolated labs, with explicit authorization. See SECURITY.md.
PRs adding skills, agents, or fixes are welcome — see CONTRIBUTING.md. Keep skills
expert-framed and token-lean; run node scripts/validate.mjs before opening a PR.
MIT — free to use, fork, and share.
Built for Claude Code. If it helped, leave a ⭐.