一个Deno脚本，能自动发现并镜像GitHub上的AI代理技能仓库。

它的作用

agent-skills-mirror 运行 GitHub 搜索查询，以查找包含 AI 代理技能、编码助手指令、Claude 技能、Copilot 指令、MCP 工具集成及相关配置文件的仓库。对于每个发现的仓库，它会执行一次 sparse checkout——只拉取与技能/提示/指令消耗相关的特定文件——并将结果写入本地 mirrors/ 目录。

为什么它存在

AI编码助手生态系统（Claude Code、GitHub Copilot、Cursor、Continue、Windsurf、MCP工具等）已在GitHub上生成了数百个高信号技能和指令仓库。本项目将它们聚合到一个地方，使得：

• 工具可以在不克隆数百个仓库的情况下消耗技能文件。
• 人类可以在一个目录树中浏览相关的指令和提示。
• 覆盖范围可以被测量并随着时间改进。

它发现了哪些类型的仓库

搜索查询针对包含以下内容的仓库：

• 特工技能 — '技能。

agent-skills-mirror

A Deno script that automatically discovers and mirrors AI agent skill repositories from GitHub.

What it does

agent-skills-mirror runs GitHub search queries to find repositories that contain AI agent skills, coding assistant instructions, Claude skills, Copilot instructions, MCP tool integrations, and related configuration files. For each discovered repository it performs a sparse checkout — pulling only the specific files relevant to skill/prompt/instruction consumption — and writes the results to a local mirrors/ directory.

Why it exists

The AI coding assistant ecosystem (Claude Code, GitHub Copilot, Cursor, Continue, Windsurf, MCP tools, etc.) has produced hundreds of high-signal skill and instruction repositories scattered across GitHub. This project aggregates them into one place so that:

• Tooling can consume skill files without cloning hundreds of repos.
• Humans can browse relevant instructions and prompts in one directory tree.
• Coverage can be measured and improved over time.

What kinds of repositories it discovers

The search queries target repositories that contain:

• Agent skills — SKILL.md, skills/**, skill/** files for Claude Code, GitHub Copilot agent mode, or custom coding agents.
• Copilot instructions — .github/instructions/**, .github/prompts/**, copilot-instructions.md.
• Claude/LLM config — CLAUDE.md, AGENTS.md, GEMINI.md, LLMs.txt.
• Cursor / Windsurf / Continue rules — .cursorrules, .cursor/rules/**, .windsurfrules, .continue/**.
• MCP integrations — .mcp/**, mcp/** directories.
• Prompt collections — prompts/**, prompt/**.

Non-GitHub sources (e.g. docs.stripe.com, open.feishu.cn, smithery.ai) are not indexed by this tool and should be consulted separately.

How to run

Prerequisites

• Deno ≥ 2.x
• Git (for sparse checkout)
• A GitHub personal access token (recommended to avoid rate-limits)

Basic run

deno run -A src/main.ts

With a GitHub token

# Using gh CLI
GH_TOKEN=$(gh auth token) deno run -A src/main.ts

# Or exporting directly
export GH_TOKEN=ghp_...
deno run -A src/main.ts

# GITHUB_TOKEN is also accepted
export GITHUB_TOKEN=ghp_...
deno run -A src/main.ts

Without a token, GitHub search is unauthenticated and heavily rate-limited (10 requests/min). With a token the limit is 30 requests/min for search.

Using the Deno task

deno task mirror

Running tests

deno task test

Controlling concurrency

MIRROR_CONCURRENCY=4 deno run -A src/main.ts

Default concurrency is 8 parallel repo checkouts.

What is mirrored and where

Mirrored files are written under mirrors/repos/@/ — for example:

mirrors/repos/vercel-labs@agent-skills/
  skills/
    vercel-v0/SKILL.md
  .github/prompts/
    ...
  manifest.json

Each directory contains: