LLM Wiki

Skill Registry

5分钟阅读

Skill Registry

Definition

定义

A skill registry is a centralized catalog service that stores, indexes, and distributes agent skills with metadata, versioning, and access control. It serves as the “npm for agent skills.”

Skill Registry 是一种集中式的目录服务，用于存储、索引和分发 Agent 技能，并提供元数据、版本控制和访问控制功能。它充当了“Agent 技能界的 npm”。

Details

详细信息

Skill registries solve the distribution and discovery problem for agent capabilities. Instead of hardcoding tools into agents, registries enable dynamic skill acquisition at runtime.

Skill registries 解决了 Agent 能力的分发与发现问题。Registry 支持在运行时动态获取 Skill，而无需将工具硬编码到 Agent 中。

Core Components

**核心组件**

1. Metadata Schema

**1. Metadata Schema**
```json { "name": "enterprise-user-management", "version": "1.2.0", "description": "Create, update, delete users in enterprise directory", "permissions": ["users:read", "users:write"], "dependencies": ["ldap-connector@^2.0.0"], "platforms": ["claude-code", "openclaw", "windsurf"], "author": "IT Department", "license": "proprietary" } ```

2. Registry API

  • GET /skills - List available skills
  • GET /skills/{name} - Get skill details
  • GET /skills/{name}/versions - List versions
  • POST /skills - Publish new skill
  • DELETE /skills/{name} - Unpublish skill
**2. Registry API** - `GET /skills` - 列出可用 skills - `GET /skills/{name}` - 获取 skill 详情 - `GET /skills/{name}/versions` - 列出版本 - `POST /skills` - 发布新 skill - `DELETE /skills/{name}` - 下架 skill

3. Access Control

  • Public registry: Open skills (agentskills.io)
  • Private registry: Enterprise-only skills
  • Role-based access: Different skills for different user roles
  • Approval workflows: Skills require review before publication
**3. 访问控制** - **Public registry**：Open skills (agentskills.io) - **Private registry**：企业专属 skills - **基于角色的访问**：针对不同用户角色提供不同的 skills - **审批流程**：Skills 发布前需经过审核

4. Versioning

  • Semantic versioning (major.minor.patch)
  • Compatibility declarations
  • Deprecation warnings
  • Automatic updates vs. pinned versions
**4. 版本管理** - 语义化版本控制 (major.minor.patch) - 兼容性声明 - 弃用警告 - 自动更新 vs. 锁定版本

Registry Types

**Registry 类型**

Public Registries

  • agentskills.io: 500+ skills, 100K+ installs, 20+ platforms
  • OpenClaw Registry: Focus on open-source skills
  • Hugging Face Spaces: Experimental skills and demos
**公共 Registry** - **agentskills.io**：500+ 技能，100K+ 安装量，支持 20+ 平台 - **OpenClaw Registry**：专注于开源技能 - **Hugging Face Spaces**：实验性技能与 Demo

Private Registries

  • Enterprise registries: Internal skills for proprietary systems
  • Compliance: Audit logs, approval workflows
  • Air-gapped: No external dependencies for secure environments
  • Examples: Alibaba Cloud ModelScope, Tencent Cloud AI Gallery
**私有 Registry** - **企业级 Registry**：面向专有系统的内部技能 - **合规性**：审计日志、审批流程 - **Air-gapped**：无外部依赖，适用于安全环境 - **示例**：Alibaba Cloud ModelScope、Tencent Cloud AI Gallery

China-Specific Considerations

  • Domestic hosting: Registries must be hosted in China for compliance
  • Content review: Skills may require government approval
  • Vendor lock-in: Platform-specific registries (DingTalk, Feishu)
  • Open source preference: Government encourages domestic open-source alternatives
**中国市场的特殊考量** - **国内托管**：为了满足合规要求，Registry 必须托管在中国境内 - **内容审查**：Skills 可能需要经过政府部门审批 - **厂商锁定**：存在特定平台的 Registry（如钉钉、飞书） - **开源偏好**：政府鼓励采用国产开源替代方案

Implementation Patterns

**实现模式**

Pattern A: Centralized Registry

  • Single source of truth
  • Easier governance and compliance
  • Single point of failure
  • Example: Alibaba Cloud ModelScope
**模式 A：集中式 Registry** - 单一事实来源 - 更易于 Governance 和 Compliance - 单点故障 - 示例：Alibaba Cloud ModelScope

Pattern B: Federated Registry

  • Multiple registries with cross-references
  • Resilience and redundancy
  • Complex synchronization
  • Example: agentskills.io + private enterprise registries
**模式 B：联邦 Registry** - 多个 Registry，具有交叉引用 - 弹性与冗余 - 复杂的同步 - 示例：agentskills.io + 私有企业 Registry

Pattern C: Embedded Registry

  • Skills bundled with agent platform
  • No external dependencies
  • Limited sharing across platforms
  • Example: Claude Code skills directory
**模式 C：嵌入式 Registry** - Skills 与 Agent 平台捆绑 - 无外部依赖 - 跨平台共享能力受限 - 示例：Claude Code Skills 目录

Key Challenges

  • Quality control: How to prevent malicious or buggy skills
  • Discovery: How users find relevant skills among thousands
  • Versioning conflicts: Dependency hell when skills require incompatible versions
  • Performance: Registry lookups add latency to agent execution
**关键挑战** - **质量控制**：如何防止恶意或有缺陷的 Skill - **发现机制**：用户如何在成千上万个 Skill 中找到相关的 Skill - **版本冲突**：当 Skill 依赖不兼容的版本时出现的依赖地狱问题 - **性能**：Registry 查询会增加 Agent 执行的延迟

Connections

连接
- 关联至：[[enterprise-agent-china/concepts/skill-factory-layers|Skill Factory Layers]], [[enterprise-agent-china/entities/agentskills-io|agentskills.io]] - 提及于：[[enterprise-agent-china/sources/skill-factory-framework|Skill Factory Framework]], [[enterprise-agent-china/sources/agentskills-io-analysis|agentskills.io Analysis]]

关系图谱

反向链接