Skip to main content
Glama

Skill-MCP 🚀

把技能当作数据的下一代 AI Agent 技能注册中心与认知规则引擎 Next-Gen AI Agent Skill Registry & Cognitive Rule Engine

License: MIT Node: >=24 MCP Protocol Tests: 91 Passing


📖 Core Philosophy & Architecture Innovation

The traditional Agent skill ecosystem faces two major industry pain points:

  1. Inefficient skill retrieval: Simple keyword matching cannot handle long-tail demand, and multi-skill orchestration lacks dependency resolution and data-flow planning capabilities.

  2. Missing rule-based governance: "Rules" such as naming conventions, code review, deep thinking, and commit conventions will almost never be proactively invoked by AI without a trigger phrase.

Skill-MCP proposes the “Dual-Track Skill Architecture”:

                               ┌───────────────────────────┐
                               │   AI Client (Claude/Cursor)│
                               └─────────────┬─────────────┘
                                             │
                                     skill_search("...")
                                             │
                                             ▼
                      ┌──────────────────────────────────────────────┐
                      │                 Skill-MCP                    │
                      ├──────────────────────┬───────────────────────┤
                      │                      │                       │
     【按需检索】      ▼                      ▼     【强制注入】       │
  ┌─────────────────────────┐         ┌────────────────────────────┐ │
  │    Tool Skills (工具型)  │         │     Rule Skills (准则型)   │ │
  ├─────────────────────────┤         ├────────────────────────────┤ │
  │ • BM25 + 向量混合召回   │         │ • 跳过搜索竞争             │ │
  │ • 多技能工作流 DAG 规划 │         │ • 随每次检索无条件跟车附带 │ │
  │ • 沙箱隔离与 HITL 权限  │         │ • 强制规范 AI 思考与代码风格│ │
  └───────────┬─────────────┘         └────────────┬───────────────┘ │
              │                                    │                 │
              └──────────────────┬─────────────────┘                 │
                                 ▼                                   │
                      ┌──────────────────────┐                       │
                      │ 统一 Payload 返回给 AI│                       │
                      └──────────────────────┘                       │
                      └──────────────────────────────────────────────┘

Related MCP server: craft

📐 Skill-MCP Skills Standard (The Skill Standard)

All skills stored in skills/ follow the unified “trinity” organization standard:

skills/<namespace>/<skill-name>/
├── skill.json         # 1. 机器契约(元数据、IO Schema、权限、执行入口、skillType)
├── SKILL.md           # 2. AI 执行 SOP 指南(角色定位、正反例、步骤规范)
└── scripts/           # 3. 可执行脚本工具(自动化校验器、数据抓取脚本、转换器等)
    └── run.ts

1. skill.json Contract Definition

{
  "schemaVersion": 1,
  "name": "naming-conventions",
  "namespace": "dev",
  "version": "1.0.0",
  "description": "代码与变量命名规范守门人:强制统一 AI 代码命名风格,约束布尔谓词、函数动词前缀与常量大写。",
  "category": "dev",
  "tags": ["naming-conventions", "code-style", "rules"],
  "triggers": ["naming", "variable", "refactor", "code"],
  "keywords": ["naming", "camelCase", "snake_case"],
  "whenToUse": "在生成、重构或审查任何编程语言的代码时必须遵守",
  "whenNotToUse": "编写纯文本说明或无代码生成的闲聊场景无需遵守",
  
  // 核心区分:'tool' (按需搜索工具) | 'rule' (永远生效的准则)
  "skillType": "rule",

  "useCases": [
    { "task": "审查并规范生成的变量与函数命名" }
  ],
  "preconditions": {},
  "io": {
    "input": { "semanticType": "text" },
    "output": { "semanticType": "text" }
  },
  "capabilities": ["dev:naming-conventions"],
  "consumes": [],
  "dependencies": [],
  "permissions": {
    "fsRead": ["*"],
    "fsWrite": [],
    "network": [],
    "tools": [],
    "env": [],
    "maxDurationMs": 5000,
    "maxCostCents": 0,
    "mutating": false
  },
  "entrypoint": {
    "kind": "inline",
    "code": "return { applied: true, rule: 'dev:naming-conventions' };"
  },
  "status": "active"
}

2. SKILL.md Authoring Standard

  • Frontmatter: Contains name, namespace, version, skillType, description.

  • Positive Specs (Do’s): List clear behavior rules.

  • Counterexample Comparison (Bad vs Good): Must provide explicit side-by-side code blocks to eliminate ambiguity for LLM understanding.


🛠️ MCP Tool Matrix (9 Tools)

工具名称

职责定位

skill_search

Hybrid search of skill types, and automatically injects global rule skills (activeRules)

skill_inspect

Inspect skill dependency topology, version conflicts, circular dependencies, and file hashes

skill_get

Get the full SKILL.md SOP manual and context instructions for a skill

skill_plan

Automatically plan multi-skill DAG data-flow, based on the topology graph and historical success recipes(Recipes)

skill_run

Execute a skill step by step in a secure sandbox (Node/Inline/Shell), controlled by the permission model (Broker)

workflow_run

Orchestrate and execute the entire workflow DAG, with two-phase human-in-the-loop authorization (HITL)

skill_feedback

Report execution feedback, triggering Thompson Sampling and Elo win-rate dynamic re-ranking

skill_register

Register new skills dynamically, with content addressing, TOFU signature binding, and quality gates

skill_stats

View historical invocation count, win rate, Elo score, and popular skill recipes


🚀 Quick Start

Prerequisites

  • Node.js: >= 24.0.0 (natively supports TypeScript type stripping and SQLite)

1. Local Run (Stdio)

Integrate directly as a local MCP service into Claude Desktop or Cursor:

# 安装依赖
npm install

# 运行测试套件(91 项自动化测试)
npm test

# 启动 Stdio MCP 服务
npm start

2. HTTP Mode for Local/Server (Streamable HTTP / SSE)

# 启动 HTTP 服务,监听 0.0.0.0:3000
node src/main.ts --http --host 0.0.0.0 --port 3000

Service-ready endpoint: http://127.0.0.1:3000/mcp (supports JSON-RPC and Server-Sent Events streaming).


🐳 Docker Deployment

# 构建镜像
docker build -t skill-mcp:latest .

# 后台运行容器(数据持久化挂载)
docker run -d \
  -p 3000:3000 \
  --name skill-mcp \
  -v $(pwd)/data:/app/data \
  skill-mcp:latest

🔌 Client Configuration Example

Claude Desktop (claude_desktop_config.json)

Local process mode (Stdio):

{
  "mcpServers": {
    "skill-mcp": {
      "command": "node",
      "args": ["/path/to/Skill-mcp/src/main.ts"]
    }
  }
}

Remote server mode (HTTP):

{
  "mcpServers": {
    "remote-skill-mcp": {
      "url": "http://your-server-ip:3000/mcp"
    }
  }
}

🛡️ Built-in Rule Skills Ecosystem (Built-in Rule Skills)

Currently built in with 13 top-level cognitive and engineering gatekeeper rule sets:

  • 🌟 Git Commit Convention: dev:git-conventional-commits

  • 🌟 Type-safety defense: dev:typescript-strict-guard

  • 🌟 Code naming gate: dev:cleaning-naming-conventions

  • 🌟 Code security review: dev:code-review-guard

  • 🌟 SOLID architecture principles: arch:clean-code-solid

  • 🌟 Enterprise-grade RESTful design: web:restful-api-standard

  • 🌟 Modern Web aesthetics: design:modern-ui-aesthetics

  • 🌟 Chain-of-verification (CoVe): reasoning:chain-of-verification

  • 🌟 Dynamic thought chain: reasoning:sequential-thinking

  • 🌟 Toyota 5-Whys: reasoning:root-cause-5whys

  • 🌟 Devil’s Advocate (adversarial): reasoning:adversarial-critic

  • 🌟 First-principles reasoning: reasoning:first-principles

  • 🌟 Multi-criteria quantitative decision-making: reasoning:decision-tradeoff-matrix


📄 Open Source License

This project is open sourced under the MIT License.

Available Tools

9 tools
skill_feedbackA

报告执行结果与评分,驱动 bandit/Elo 排名与备选检索(学习闭环;会覆盖同一次 skill_run 的自动记账,不重复计数)。

Inputs: skill (required), outcome (required), task, runId, +5 more.

skill_getC

获取 skill 的完整包:manifest + SKILL.md 正文 + 文件清单 + 已解析依赖(渐进披露的正文环节)。

Inputs: key, name.

skill_inspectC

查看 skill 的 manifest、依赖树与权限清单(不含正文,成本低)。

Inputs: key, name.

skill_planC

把任务规划成 skill 组合的 workflow DAG(分解-检索-组合;无 LLM 时用能力闭包图搜索)。

Inputs: task (required), context.

skill_registerA

注册或升级 skill(append-only;v3:trust 模式需 Ed25519 签名,gate 模式下升级需自带 tests 过沙箱)。

Inputs: manifest (required), files, signature, publisherKey.

skill_runA

在权限沙箱中执行单个 skill。权限不足时返回 requiresApproval(两段式 HITL,approveAsks: true 授权后重试)。

Inputs: skill (required), task, inputs, timeoutMs, +2 more.

skill_statsA

查看 skill 的评价指标:成功率(Wilson)、Elo、质量分、延迟成本,以及挖掘出的 workflow recipe。

Inputs: key.

workflow_runB

执行 skill_plan 产出的 DAG:拓扑序数据流、重试、检查点、整链授权。

Inputs: dag (required), task, inputs, approveAsks, +1 more.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.7/5.0

    Scored across 9 tools

    Disambiguation5/5

    每个工具对应一个明确的职责阶段:搜索、查看元信息、获取全文、规划、执行单技能、执行工作流、反馈、注册、查看统计,彼此边界清晰。workflow_run 与 skill_run 通过单个体与 DAG 执行明确区分,不易产生误选。

    Naming Consistency4/5

    大多数工具遵循 skill_<action> 的模式,如 skill_search、skill_inspect、skill_register,具有较高的可预测性。但 workflow_run 没有使用 skill_ 前缀,且 skill_stats、skill_feedback 以名词而非动词结尾,存在少量风格不一致。

    Tool Count5/5

    9 个工具覆盖了从发现、规划、执行到反馈、学习、注册、查询的完整闭环,每个工具都有独立且必要的职责。数量处于合理区间,没有冗余或缺失。

    Completeness5/5

    工具面覆盖了 skill 的生命周期:注册/升级、检索、检查、获取、单技能执行、工作流执行、反馈学习、统计评估。缺少删除/下架属于 append-only 设计下的有意限制,不构成明显缺口。

    Maintenance

    ActivityInactive
    ResponsivenessOutdated data

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      F
      maintenance
      Enables AI agents to intelligently match tasks to skills through semantic embeddings, track skill effectiveness, detect skill gaps, and discover new skills from external sources.
      Apache 2.0
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables AI agents to index and search local skill libraries, persist and resume task checkpoints, and run validated workflow plans with controlled approval and write permissions.
      MIT
    • A
      license
      Not graded
      quality
      C
      maintenance
      Enables AI agents to search, retrieve, score, compose, and verify reusable agent skills from the SkillForge registry via MCP.
      21 npm
      MIT