SuperSkill
Runtime intelligence layer for AI agent skills that scores, resolves, and loads the best skill for a task, managing context and collision resolution.
README
SuperSkill
The runtime intelligence layer for AI agent skills.
The Problem
skills.sh solved skill discovery and installation — 90,000+ skills, 40+ agents, one standard. But installation is static. You browse, pick, install, and hope you chose right.
When you have 15 skills installed and say "write tests for my Go API" — which one loads? What about the three competing TDD skills? What if the model's context window can't fit them all?
skills.sh is where you find skills. SuperSkill is what loads the right one at the right time.
How It Works
- Install skills however you want —
npx skills add, manual download, or use the 87 built-in skills - Describe your task — "write tests for my Go API"
- SuperSkill scores, resolves, and loads — trigger-based matching picks the best skill, resolves collisions, respects your context budget
SuperSkill scans your installed skills (from ~/.claude/skills/, ~/.cursor/skills/, etc.), merges them with its built-in catalog, and routes tasks to the best match at runtime.
What SuperSkill Adds
| Layer | What it does | Why you need it |
|---|---|---|
| Router | Task description → best matching skill | No manual browsing or memorizing skill names |
| Arbiter | 3 TDD skills installed? Profile picks the winner | Prevents collision chaos and token waste |
| Scanner | Discovers installed skills from any source | Works with skills.sh, manual installs, or built-in |
| Memory | Remembers activations within a session | No redundant fetches for similar tasks |
| Vault | ADRs, learnings, sessions, project context | Persistent knowledge that survives across sessions |
Quick Start
Claude Code Plugin (recommended)
/plugin marketplace add permanu/superskill
/plugin install superskill
Any MCP-compatible tool
npm install -g superskill
Then add as an MCP server:
# Claude Code
claude mcp add superskill -e VAULT_PATH=~/Vaults/ai -- npx -y superskill
# Cursor, Claude Desktop, Codex, Gemini CLI — add to MCP config:
{
"mcpServers": {
"superskill": {
"command": "npx",
"args": ["-y", "superskill"],
"env": { "VAULT_PATH": "~/Vaults/ai" }
}
}
}
Works with skills.sh
Install skills from the ecosystem, SuperSkill routes to them automatically:
npx skills add anthropics/skills # Anthropic's official skills
npx skills add affaan-m/everything-claude-code # ECC skill suite
# SuperSkill discovers and scores these on next activation
Supported AI Tools
| Tool | Setup | Status |
|---|---|---|
| Claude Code | Plugin or MCP | Verified |
| Claude Desktop | MCP config | Verified |
| Cursor | MCP config | Verified |
| Codex CLI | MCP config | Verified |
| Gemini CLI | MCP config | Verified |
| OpenCode | MCP config | Community |
| Crush CLI | MCP config | Community |
| Droid | MCP config | Community |
| Windsurf | MCP config | Community |
| Aider | MCP config | Community |
| Continue | MCP config | Community |
Built-in Skill Catalog
<details> <summary>87 skills across 28 domains from 9 repos (click to expand)</summary>
Core Workflow — loaded for every project:
| Domain | Skills | Description |
|---|---|---|
| TDD | 8 | Red-green-refactor, Go/Python/Django/Spring/C++ testing, E2E |
| Planning | 3 | Implementation planning, CEO/eng review, execution |
| Code Review | 4 | PR review, Go review, Python review, feedback workflow |
| Debugging | 2 | Systematic debugging, investigation |
| Verification | 4 | Build/lint/type gates, Django/Spring verification |
| Brainstorming | 2 | Structured ideation, office hours |
| Agent Orchestration | 5 | Autonomous loops, RFC pipelines, subagents, parallel dispatch |
| Security | 5 | OWASP review, AgentShield scanning, Django/Spring security |
| Shipping | 2 | CI/CD, deployment patterns |
| Frontend Design | 5 | Anthropic official, Design Taste, Bencium UX, FDP, UI/UX Pro Max |
| Git Workflow | 2 | Worktrees, branch management |
Language & Framework — loaded when your stack matches:
| Domain | Skills | Description |
|---|---|---|
| Go | 2 | Idiomatic patterns, conventions |
| Python | 2 | Pythonic idioms, PEP 8 |
| Django | 3 | Architecture, DRF, ORM, security, TDD |
| Spring Boot | 4 | Architecture, security, TDD, verification |
| Swift | 4 | SwiftUI, concurrency, actors, protocol DI |
| C++ | 2 | Core Guidelines, GoogleTest |
| Java | 2 | Standards, JPA/Hibernate |
| Database | 3 | PostgreSQL, migrations, ClickHouse |
| Docker | 1 | Compose, container security |
Infrastructure & Patterns:
| Domain | Skills | Description |
|---|---|---|
| API Design | 1 | REST patterns, pagination, versioning |
| Frontend Patterns | 1 | React/Next.js state and performance |
| Backend Patterns | 1 | Node/Express server patterns |
| Coding Standards | 1 | Universal TS/JS/React standards |
Specialized:
| Domain | Skills | Description |
|---|---|---|
| Content & Business | 6 | Articles, investor materials, outreach, market research |
| 3D Animation | 5 | Three.js, GSAP, React Three Fiber, Framer Motion, Babylon.js |
| Agent Engineering | 4 | Agent harness, eval, cost optimization |
| Meta/Tooling | 5 | Skill management, compaction, learning, browsing |
</details>
MCP Tools
| Tool | Description |
|---|---|
superskill |
Route to the best skill by task description, domain, or skill ID |
vault_skill |
Skill catalog, collisions, resolution, generation |
vault_project_context |
Auto-detected project context from CWD |
vault_init |
Generate draft context.md from a git repo |
vault_task |
Task management (add/list/update/board) |
vault_decide |
Log architecture decisions |
vault_learn |
Capture and list learnings |
vault_resume |
Resume context — recent sessions, interrupted work, next steps |
vault_session |
Multi-agent session coordination |
vault_read |
Read file or directory from vault |
vault_write |
Write/append/prepend content |
vault_search |
Full-text search across vault |
vault_prune |
Archive stale content |
vault_stats |
Vault content statistics |
vault_deprecate |
Mark items as deprecated |
Configuration
| Variable | Default | Description |
|---|---|---|
VAULT_PATH |
~/Vaults/ai |
Path to knowledge vault |
MAX_INJECT_TOKENS |
1500 |
Max tokens for context injection |
SESSION_TTL_HOURS |
2 |
Session heartbeat TTL |
Roadmap
See GitHub Milestones for planned work.
Contributing
See CONTRIBUTING.md.
License
AGPL-3.0-or-later — See LICENSE
Copyright 2026 Permanu (Atharva Pandey)
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。