repo-evals · all evaluated reposrepo-evals · 所有评测过的仓库
All Evaluated Skills & Repos已评测的所有 skill 和仓库
Master index of every repo evaluated under repo-evals. Click into any row for the full bilingual dossier — claims, evidence, score breakdown, deployment notes, and the full benefits block (who / when / without / with). Sort by clicking a column header; filter by category or search by name. repo-evals 评测过的所有仓库总目录。点任意一行进入完整双语 dossier —— claim、证据、分数明细、部署成本、完整 benefits 块（谁 / 什么时候 / 没它 / 有它）。点列头排序；按类别过滤或搜名字。
Total总数
41
🏭 Production可用于生产
3
🛠 Available可使用
32
⚠️ Risky有风险
6
🛑 Don't use不可使用
0
By domain:按业务大类:
I want to …:我想干嘛?
| # | Repo仓库 | What it does做什么 | When you'd use it什么时候用上 | Cost & deps成本 / 依赖 | Score分数 | Dossier详细 |
|---|---|---|---|---|---|---|
| 1 | geekjourneyx/md2wechat-skill ↗ | A Go CLI that converts Markdown to WeChat-editor-ready HTML in one command — 228-test polished support layer, 5 discovery commands returning consistent JSON, multi-provider LLM hooks for "humanize" + "write from idea" modes.一个 Go CLI,把 Markdown 一行命令转成微信公众号编辑器可用的 HTML —— 228 个测试覆盖的成熟支撑层,5 个 discovery 命令统一返回 JSON,可选多家 LLM 接入"humanize" + "从想法直接写"两种模式。 | You finished an article in Markdown — code blocks, images, lists, footnotes. WeChat editor doesn't accept Markdown. You need clean styled HTML you can paste into the editor and hit publish.你写完一篇 Markdown(代码块、图片、列表、脚注都齐),但是公众号编辑器不认 Markdown。你需要一份干净有调性的 HTML,贴进编辑器一键发布。 | MIT-licensed CLI is free. Core conversion calls md2wechat.cn (free tier exists; check for API key requirement). LLM modes (humanize / write from idea) need YOUR provider API key (gemini / modelscope / openai / openrouter / tuzi / volcengine). Build from source needs Go 1.26.1+; precompiled via Homebrew / npm / install script.MIT 许可的 CLI 免费。核心转换调 md2wechat.cn(有免费档,API key 视情况而定)。LLM 模式(humanize / 从想法直接写)需要你自己的 provider API key(gemini / modelscope / openai / openrouter / tuzi / volcengine)。从源码编译需要 Go 1.26.1+;Homebrew / npm / 安装脚本可拿预编译版本。 | 91 🏭 Production-ready可用于生产 | → |
| 2 | heygen-com/HyperFrames ↗ | Write HTML, render video — AI agents do the authoring, HyperFrames does the rendering. Open-source (Apache-2.0) framework from HeyGen with a 13-skill bundle that lets Claude/Cursor/Codex produce real MP4s instead of just storyboards.用 HTML 写视频，让 AI 智能体动手做 — HeyGen 出品的开源视频框架（Apache-2.0），随包 13 个 skill 让 Claude/Cursor/Codex 不再只生成 storyboard，而是直出可以发布的 MP4。 | You give an agent a prompt like 'a 10-second product intro with fade-in title, background video, and music'. Agent writes an HTML composition with data attributes, calls `hyperframes render`, hands back an MP4. No bundler step, no proprietary DSL, no per-render fee.你给 agent 一句话：「做个 10 秒产品片头，标题淡入，背景视频，配 BGM」。Agent 写一份带 data 属性的 HTML 组合，调 `hyperframes render`，把 MP4 还给你。零打包步骤、零私有 DSL、零按次计费。 | Free, open source, Apache-2.0. Self-hosted: pay your own CPU + FFmpeg time. No SaaS gate, no per-render fee, no seat cap, no company-size threshold. Skills install layer routes through `vercel-labs/skills` (also free).免费开源，Apache-2.0。自托管：自己出 CPU 和 FFmpeg 的电费。不锁 SaaS，不按次计费，不限座位数，不卡公司规模。Skills 走 `vercel-labs/skills`（也免费）。 | 89 🏭 Production-ready可用于生产 | → |
| 3 | anthropics/skill-creator ↗ | Anthropic's official Claude Code skill for creating, evaluating, and iteratively improving skills — ships a full LLM-output eval harness (3 grader agents, blind-comparison, benchmark aggregation, HTML eval-viewer) so quality measurement is built into the workflow, not an afterthought.Anthropic 官方的「写 skill 的 skill」：创建 / 评测 / 迭代改进 skill 一站式。自带完整的 LLM 输出评测工具链（3 个评测 agent、盲对比、benchmark 聚合、HTML 评测看板），把质量测量做成工作流的一部分，不是事后补的。 | You're about to ship a new skill (or a v2 of an old one). You don't want to find out it's worse than v1 from user complaints — you want to A/B test now, with quantitative evidence, against the old version on a benchmark of real prompts.你准备发一个新 skill(或者老 skill 的 v2)。你不想等用户抱怨才发现 v2 比 v1 还差 —— 你想现在就 A/B 测,用量化证据,在一组真实 prompt 上跟老版本对比。 | Apache 2.0, free skill itself. Cost is on the API side — every eval round multiplies LLM calls across grader + comparator + analyzer sub-agents. Budget: a 20-prompt benchmark with 3 graders + 1 comparator over 5 iterations is ~300 LLM calls. Setup labour: clone anthropics/skills (a catalog repo) and copy skill-creator/ into your skills dir; not a one-line install.Apache 2.0,skill 本身免费。成本在 API 侧 —— 每一轮评测会多发 LLM 调用(grader + comparator + analyzer 三个子 agent)。预算: 20 条 prompt 的 benchmark × 3 grader + 1 comparator × 5 轮 ≈ 300 次 LLM 调用。安装手工成本: clone 整个 anthropics/skills(catalog 仓库)再把 skill-creator/ 拷到你的 skills 目录;不是一行命令。 | 81 🏭 Production-ready可用于生产 | → |
| 4 | Panniantong/Agent-Reach ↗ | Give your AI agent eyes to see the entire internet — one CLI to read & search Twitter / Reddit / YouTube / GitHub / Bilibili / 小红书. MIT-licensed, zero API fees, 19K stars.给你的 AI agent 装上看互联网的眼睛 —— 一个 CLI 读 + 搜 Twitter / Reddit / YouTube / GitHub / B 站 / 小红书。MIT 协议,无 API 费,19K stars。 | You ask your AI: "What are people saying about <topic> on X / Reddit / YouTube right now?". Without a tool like this the agent fabricates or refuses. With Agent-Reach the agent runs real searches and returns sourced summaries.你问 AI: "现在 X / Reddit / YouTube 上大家在怎么聊 <某话题>?"。没有这种工具,agent 要么编要么拒绝。有了 Agent-Reach,agent 真去搜,拿带来源的总结回来。 | MIT, free. No API fee tier needed. Cost is anti-bot drift per platform — Twitter / X especially churns post-2023, so durability not stability. Your agent's normal LLM cost still applies.MIT 免费。不要付费档 API。成本是各平台反爬漂移 —— 尤其 Twitter / X 在 2023 年后 churn 很猛,看持久性不是稳定性。你 agent 平常的 LLM 成本照算。 | 79 🛠 Available可使用 | → |
| 5 | router-for-me/CLIProxyAPI ↗ | A Go proxy server that wraps your Gemini CLI / Antigravity / ChatGPT Codex / Claude Code OAuth sessions as OpenAI / Gemini / Claude / Codex compatible HTTP APIs — multi-account round-robin, function calling, multimodal, streaming, and a reusable Go SDK to embed it. 8-platform binaries, MIT, very actively developed.一个 Go 代理服务器：把你已经登录的 Gemini CLI / Antigravity / ChatGPT Codex / Claude Code 会话包装成 OpenAI / Gemini / Claude / Codex 兼容的 HTTP API。多账号轮询、函数调用、多模态、流式、还有可嵌入的 Go SDK。8 平台二进制，MIT，更新非常活跃。 | You have a working CLI subscription. Your other code wants to call "Claude" but expects the OpenAI API shape. Or your token expired. Or you want round-robin across 3 personal accounts so heavy work doesn't burn one quota. You don't want to write a wrapper for each provider.你有可用的 CLI 订阅。但你别的代码想调"Claude",而它期待 OpenAI API 形状。或者 token 过期了。或者你想 3 个个人账号轮询,不让重活把一个配额烧光。你不想给每家 provider 写一遍 wrapper。 | MIT, free. Reuses CLI subscriptions you already pay for — no new API keys for the wrapped providers. Optional providers (Kimi / Vertex) need their own keys. No external services required at runtime. Token storage at `./auths/` — keep local. Self-hosted, no telemetry.MIT 免费。复用你已经在付的 CLI 订阅 —— 包装的 provider 不需要新 API key。可选的 Kimi / Vertex 走它们自己的 key。运行时不需要外部服务。Token 存 `./auths/` —— 本地保管。自托管 / 无遥测。 | 78 🛠 Available可使用 | → |
| 6 | zinan92/repo-evals ↗ | A claim-first evaluation framework that turns "is this skill / repo any good?" into a 0-100 score with a bilingual dossier — auditable point-by-point, comparable to the 30 other repos already evaluated under the same model.一个 claim-first 的评测框架,把"这个 skill / repo 到底能不能用?"变成一份带 0-100 分的双语 dossier —— 每一分可以追到证据,可以跟我们已经评测过的 30 个 repo 横向对比。 | You're about to commit budget / install / brand to a new tool — a crawler library for a research project, a methodology bundle for your team's coding agent, a slide-generator for next month's workshop. You want a structured "is this real?" check, not a vibe.你准备投入(预算 / 安装 / 品牌)到一个新工具 —— 给研究项目挑爬虫库、给团队 coding agent 挑方法论合集、给下个月 workshop 挑 deck 生成器。你想要一份结构化的"它是真的吗?"检查,不只是凭感觉。 | Free framework but unlicensed (README claims MIT, no LICENSE file yet). Requires Python 3.11+ + PyYAML. Manual labour cost: writing claim-map.yaml for each new repo (~30-60 min for a thoughtful 6-10 claim set). No external services, no API keys.框架免费但 license 缺失(README 写 MIT 但没 LICENSE 文件)。需要 Python 3.11+ + PyYAML。人工成本: 每个新 repo 要手写一份 claim-map.yaml(认真写 6-10 条 claim 大约 30-60 分钟)。没有外部服务,没有 API key。 | 78 🛠 Available可使用 | → |
| 7 | obra/superpowers ↗ | A 14-skill methodology bundle that auto-triggers when your coding agent starts work — brainstorming, planning, TDD, subagent dispatch, debugging, code review, git worktrees.给 coding agent 的 14 个方法论 skill 合集，agent 一开工就自动触发 —— brainstorming / 写计划 / TDD / 子 agent 拆活 / 系统性 debug / 代码 review / git worktrees。 | You're about to ask your agent to build something non-trivial — a new auth flow, a DB migration, a refactor that touches 5 files. You want it to think it through first, not rush to a diff.你要让 agent 做一个不是 5 行能搞定的活 —— 新加一套登录流程、改库表、重构跨 5 个文件的模块。你希望它先想清楚再动手,不是上来就直接 diff。 | Free skill (MIT). Cost is on the agent side — 14 skills auto-trigger means more tokens consumed per task before any code gets written. Need an existing agent harness (Claude Code / Codex / Cursor / Gemini CLI / OpenCode / Factory Droid / GitHub Copilot CLI).skill 本身免费(MIT)。成本在 agent 这边 —— 14 个 skill 自动触发,每个任务在写代码前要先消耗更多 token。需要已有的 agent 平台(Claude Code / Codex / Cursor / Gemini CLI / OpenCode / Factory Droid / GitHub Copilot CLI)。 | 77 🛠 Available可使用 | → |
| 8 | iamzhihuix/skills-manage ↗ | Manage skills across 28 AI coding tools (Claude Code, Cursor, Codex, Gemini CLI, Trae, etc.) from one desktop app. Edit a skill once, sync to every tool.一个桌面应用，帮你管理 Claude Code、Cursor、Codex、Gemini CLI、Trae 等 28 个 AI 编程工具的技能库。技能改一处，所有工具同步更新。 | You wrote a useful skill (or installed someone else's). You want it active in every tool you use. You don't want to maintain 5 copies, hand-update each one when you tweak the SKILL.md, or forget which copy is the latest.你写了个好用的 skill(或者装了别人的)。你想让它在所有用到的工具里都生效。你不想维护 5 份副本,每次改 SKILL.md 都手动同步,或者搞不清哪份是最新的。 | Free desktop app (MIT). macOS DMG is unsigned at v0.9.1 — first launch needs `xattr -dr com.apple.quarantine ~/Applications/skills-manage.app`. No external services; doesn't talk to upstream LLM providers itself.免费桌面 app(MIT)。v0.9.1 的 macOS DMG 未签名 —— 首次启动需要手动跑 `xattr -dr com.apple.quarantine ~/Applications/skills-manage.app`。不调外部服务,不跟上游 LLM provider 通信。 | 77 🛠 Available可使用 | → |
| 9 | brokermr810/QuantDinger ↗ | A self-hosted Docker Compose stack for AI-assisted quantitative trading: charting, indicator and strategy generation, backtesting, and live execution across crypto / IBKR stocks / MT5 forex — talks to your AI agent through an MCP server, brings its own Postgres + Redis + Flask backend + React frontend.一个自托管的 AI 量化交易 Docker Compose 平台：行情图表、AI 辅助生成指标和策略、回测、加密货币 / IBKR 美股 / MT5 外汇实盘下单都在一套 stack 里。通过 MCP server 接你的 AI Agent，自带 Postgres + Redis + Flask 后端 + React 前端。 | You have an idea ("buy on RSI dips, exit on MA20 cross"), want to test it on 3 years of data, iterate the rules, then put a small live position. You want the AI to write the indicator code + the strategy code so you don't hand-code each variation, but you want the keys + the server in your control.你有想法("RSI 回调买入,MA20 上穿离场"),想用 3 年数据回测,迭代规则,然后挂一个小仓位实盘。你希望 AI 替你写 indicator 代码 + 策略代码省得每次手写,但 key + 服务器在你自己手里。 | MIT, free OSS. You pay for: (1) one of LLM providers (OpenAI / DeepSeek / Grok) for code generation; (2) broker accounts (IBKR / MT5 / crypto) you'd already have. AWS Marketplace AMI exists if you don't want to manage Docker yourself (paid).MIT 免费开源。你付: (1) LLM provider(OpenAI / DeepSeek / Grok)三选一,用于代码生成;(2) 已有的 broker 账号(IBKR / MT5 / 加密所)。不想自己管 Docker 可买 AWS Marketplace AMI(付费)。 | 76 🛠 Available可使用 | → |
| 10 | NanmiCoder/MediaCrawler ↗ | One Python CLI that crawls posts / comments / creators / search from 7 Chinese platforms (Xiaohongshu, Douyin, Kuaishou, Bilibili, Weibo, Tieba, Zhihu) into 8 storage backends.一个 Python CLI,把 7 个中文平台(小红书 / 抖音 / 快手 / B 站 / 微博 / 贴吧 / 知乎)的帖子 / 评论 / 创作者 / 搜索抓进 8 种存储后端。 | You want this week's top Xiaohongshu posts under a tag, or all comments on a Douyin video, or a creator's last 200 Weibo posts. You'll log in with your real account once (QR / phone / cookie), let it run on your laptop, and dump to CSV / SQLite for downstream analysis.你想要本周小红书某个 tag 下的爆款,或者某个抖音视频的全量评论,或者某创作者在微博的最近 200 条帖子。准备真实账号登一次(扫码 / 短信 / cookie),挂在自己电脑上跑,导出 CSV / SQLite 给下游分析用。 | OSS but custom non-commercial license — research / personal only. Need a real account on each platform you want to crawl. Will break when platforms update DOM / signing — budget ongoing maintenance time. Paid PRO version exists (removes Playwright, adds resume / multi-account); not the same product.开源但自定义非商用 license —— 只能用于研究 / 个人。每个想抓的平台需要真实账号。平台改版会导致代码失效,要预留持续维护时间。有付费 PRO 版本(去掉 Playwright,加断点续传 / 多账号),不是同一个东西。 | 75 🛠 Available可使用 | → |
| 11 | THU-MAIC/OpenMAIC ↗ | An open-source Next.js + LangGraph multi-agent platform that turns any topic or document into an interactive classroom — AI teachers + AI peers + auto-generated slides / quizzes / interactive simulations / whiteboard / TTS, with 5 LLM providers and 5 TTS providers wired in. Vercel-deploy or self-host; OpenClaw skill ships for chat-app integration.清华团队开源的 Next.js + LangGraph 多 Agent 教学平台：把任何题目或文档变成互动课堂 —— AI 老师 + AI 同学一起讲、画白板、生成幻灯片 / 测验 / 互动模拟 / TTS 朗读。配套 5 家 LLM provider + 5 家 TTS provider，可一键 Vercel 部署或本机自建；自带 OpenClaw skill，能从飞书 / Slack / Discord / Telegram 直接生成课堂。 | You have a 30-page paper on a topic you don't fully understand. Reading it cover-to-cover takes 2 hours and you'll forget half. You want a 20-minute classroom where AI teachers walk you through the key points, drawing on a whiteboard, with quiz checkpoints, and an AI peer asking the questions you wouldn't have thought to ask.你有一份 30 页的陌生主题论文。完整读一遍要 2 小时,读完忘一半。你想要 20 分钟的课堂,AI 老师在白板上讲核心点,带 quiz 检查,加一个 AI 同学问你想不到要问的问题。 | AGPL-3.0 — any modified deployment must release source. You bring your own LLM key (5 supported: OpenAI / Anthropic / Google / DeepSeek / Grok) and optional TTS key (5 supported: OpenAI / Azure / GLM / Qwen / MiniMax). Per-classroom cost is non-trivial when TTS is on — budget 1-3 USD per 20-minute class depending on model + voice.AGPL-3.0 —— 任何二次部署改了代码必须开源。自己带 LLM key(5 家可选: OpenAI / Anthropic / Google / DeepSeek / Grok),TTS key 可选(5 家: OpenAI / Azure / GLM / Qwen / MiniMax)。开 TTS 的话每堂课成本不低 —— 预算每 20 分钟课 1-3 USD,看模型 + 声音。 | 75 🛠 Available可使用 | → |
| 12 | op7418/Humanizer-zh ↗ | A Claude Code skill that rewrites Chinese AI-generated text to sound human — removes 24 AI-writing tells (三段式法则, 否定式排比, AI 词汇, 破折号滥用) and scores the result on a 50-point rubric.一个 Claude Code 技能，把中文 AI 生成文本改写得像人说的话 — 识别并修掉 24 种 AI 味（三段式法则 / 否定式排比 / AI 词汇 / 破折号滥用），改完还给 1-50 分打分。 | You have an 800-2000-char Chinese draft. Bones are right but the prose reads like a machine wrote it. You need it human-sounding before publishing — without losing the structure or argument.你手里一份 800-2000 字的中文初稿。骨架对的,但读起来像机器写的。发出去前要让它"像人说的话",同时不丢结构和观点。 | Free skill (MIT). Needs Claude Code subscription. `npx skills add` install depends on an external CLI; fall back to `git clone` into ~/.claude/skills/humanizer-zh if that fails. No external API beyond Claude.skill 免费(MIT)。需要 Claude Code 订阅。`npx skills add` 依赖外部 CLI,跑不通就 `git clone` 到 ~/.claude/skills/humanizer-zh。除 Claude 外不需要其他 API。 | 75 🛠 Available可使用 | → |
| 13 | HughYau/qiushi-skill ↗ | 10 dialectical-materialism methodology skills (contradiction analysis, investigation-first, mass-line, criticism, protracted strategy, etc.) packaged as installable AI agent skills for Claude Code / Codex / Cursor / Hermes / NanoBot / OpenClaw / OpenCode. Each skill grounds itself in classical-text excerpts. MIT, multilingual (中文 + EN).10 个唯物辩证法 + 实践哲学方法论 AI agent skill（矛盾分析、调查在先、群众路线、批评与自我批评、持久战、集中兵力 ……），适配 Claude Code / Codex / Cursor / Hermes / NanoBot / OpenClaw / OpenCode 7 个平台。每条方法论都引用经典著作原文。MIT，中英双语。 | You're about to ask the agent to do something hard — analyse a complex business decision, investigate an ambiguous bug, plan a multi-month strategy. You want it to push back when your framing is wrong, dig before answering, and stick with the problem when the easy answer fails.你准备让 agent 做难事 —— 分析一个复杂商业决策、调查一个模糊 bug、规划几个月的战略。你希望它在你的提问框架不对时反问、给答案前先挖一下、容易答案失败时不放弃。 | Free skill (MIT). `npx qiushi-skill install <platform>` installs into one of 7 supported agent platforms. No API keys beyond your existing agent's. Pure markdown skills — no extra LLM calls beyond what your agent normally does.skill 免费(MIT)。`npx qiushi-skill install <platform>` 装到 7 家支持的 agent 平台之一。除已有 agent 外无需 API key。纯 markdown skill —— 不会比你 agent 平常多调 LLM。 | 73 🛠 Available可使用 | → |
| 14 | wechat-article/wechat-article-exporter ↗ | An online + self-hostable Nuxt + Vue web app for batch downloading WeChat 公众号 articles — HTML format preserves layout 100%, exports view counts + comment data, three deployment paths (hosted / Docker / Cloudflare).一个在线 + 可自托管的 Nuxt + Vue Web app,批量下载微信公众号文章 —— HTML 格式 100% 还原排版,导出阅读量 + 评论数据,三种部署(托管 / Docker / Cloudflare)。 | You follow some WeChat 公众号 you want to archive — either to study writing style, to back up your own posts, or to research competitor content. WeChat doesn't offer bulk export. You want one-click "download all this account's articles to local HTML / markdown".你关注的某些公众号要归档 —— 研究写作风格 / 备份自己的稿子 / 调研竞品。微信不给批量导出。你要一键"把这个号所有文章下到本地 HTML / markdown"。 | MIT, free. Hosted site is free; self-hosting needs Docker or Cloudflare Pages (free tier sufficient). No API keys to bring. 公众号 access does require login (cookie / WeChat-token); see project docs for current auth flow.MIT 免费。托管站点免费;自托管要 Docker 或 Cloudflare Pages(免费档够用)。不需要带 API key。访问公众号需要登录(cookie / WeChat token);最新鉴权流程看项目文档。 | 70 🛠 Available可使用 | → |
| 15 | zinan92/park-intel ↗ | A self-hosted FastAPI pipeline that crawls 10+ news/social sources (RSS, HN, Reddit, GitHub, Xueqiu, etc.), tags articles, clusters them into narrative events with signal scores, and serves it all as a REST API + React feed on :8001.一个自建的 FastAPI 情报采集管道 — 从 10+ 新闻和社交源（RSS、HN、Reddit、GitHub、雪球等）抓文章、打标签、聚类成事件并打分，最后通过 :8001 的 REST API + React feed 呈现。 | You want to wake up and immediately see what stories are clustering across HN + Reddit + RSS + Twitter, with relevance scores against your watchlist tickers. You don't want the noise of raw RSS readers; you want events with multi-source coverage.早上起来想直接看到 HN + Reddit + RSS + Twitter 在聚合什么故事，结合你的 watchlist 给出相关性打分。不想要 RSS 阅读器那种原始噪音，要的是多源验证过的事件。 | Free as OSS (no LICENSE file though). Need Python 3.11+, SQLite (built-in), Node.js for frontend. Optional: ANTHROPIC_API_KEY for LLM relevance scoring (~$10/mo for personal use).开源免费（虽然还没 LICENSE 文件）。要 Python 3.11+、SQLite（内置）、Node.js 跑前端。可选：ANTHROPIC_API_KEY 用于 LLM 相关性打分（个人用大约月 $10）。 | 70 🛠 Available可使用 | → |
| 16 | Usagi-org/ai-goofish-monitor ↗ | A self-hosted FastAPI + Vue dashboard for monitoring 闲鱼 (XianYu, China's biggest secondhand market) — natural-language criteria, multi-task scheduling, Playwright scraping, multi-modal LLM filtering, and 6 notification channels (ntfy/Bark/WeWork/Telegram/Gotify/Webhook). One docker-compose, SQLite, Chromium bundled.一个自建的 FastAPI + Vue 闲鱼监控看板：用自然语言写「我想要什么」，多任务并发跑 Playwright 抓页面 + 多模态 LLM 筛选，命中就推 6 种渠道（ntfy / Bark / 企微 / Telegram / Gotify / Webhook）。一行 docker compose，SQLite，镜像自带 Chromium。 | You're hunting a specific item — say a Sony A7R III in good condition under ¥8000, or a 16-inch MacBook Pro M2 Max with battery cycles under 200. New listings appear on XianYu every minute. You can't refresh the app every 10 minutes for weeks.你在蹲特定货 —— 比如成色不错的 Sony A7R III 8K 以下,或者 16 寸 MacBook Pro M2 Max 循环数 200 以下。闲鱼上每分钟有新挂牌。你不可能每 10 分钟手动刷一次,刷几周。 | MIT, free. You pay LLM costs (default OpenAI-compatible endpoint is modelscope.cn — China-friendly but data goes to a 3rd-party gateway; change `OPENAI_BASE_URL` if you want your own). Real XianYu account needed (cookie from companion Chrome extension). Self-hosted; Docker required.MIT 免费。你付 LLM 费(默认 OpenAI 兼容端点是 modelscope.cn —— 国内友好但数据过第三方网关;介意就换 `OPENAI_BASE_URL` 到自己 endpoint)。需要真实闲鱼账号(用配套 Chrome 扩展导 cookie)。自托管,要 Docker。 | 69 🛠 Available可使用 | → |
| 17 | AIDC-AI/Pixelle-Video ↗ | An Apache 2.0 fully-automated short-video engine — give it an idea, get back a finished MP4. Multi-LLM provider abstraction, Docker + Windows distribution, FastAPI backend, mkdocs site. 11.5K stars, AIDC-AI (Alibaba) lab.Apache 2.0 全自动短视频引擎 —— 给个想法,拿到成品 MP4。多 LLM provider 抽象,Docker + Windows 分发,FastAPI 后端,mkdocs 站点。11.5K stars,AIDC-AI(阿里)实验室出品。 | You have an idea (or 100 ideas in a queue). You want each idea turned into a 30-90 second short video — script, voiceover, scene composition, music, finished MP4. Manual production is 1-2 hours per clip; you want it down to 10 minutes mostly-unattended.你有一个(或 100 个排队的)想法。你想每个想法变成 30-90 秒短视频 —— 脚本 / 配音 / 场景 / 音乐 / 成品 MP4。手工生产一条 1-2 小时;你想压到 10 分钟基本无人值守。 | Apache 2.0, free OSS. You bring LLM provider keys (OpenAI / Anthropic / local — your choice via llm_presets) + asset / video provider keys (depending on which scene/voiceover services configured). Per-video LLM + asset cost depends on quality settings — budget 0.5-3 USD per finished short clip.Apache 2.0 免费开源。自带 LLM provider key(OpenAI / Anthropic / 本地 —— 通过 llm_presets 选)+ 资产 / 视频 provider key(看你配的场景 / 配音服务)。每条视频 LLM + 素材成本看质量设置 —— 一条成品预算 0.5-3 美元。 | 69 🛠 Available可使用 | → |
| 18 | remotion-dev/@remotion/skills ↗ | The official @remotion/skills package — one Claude/Codex/Cursor skill (`remotion-best-practices`, 340 lines) plus 35 deep-dive markdown rules covering audio, captions, ffmpeg, fonts, lottie, mapbox, and more. Lives inside the Remotion monorepo; not published on npm; consumed by Remotion's own tooling.Remotion 官方 @remotion/skills 包：一个 Claude/Codex/Cursor skill（remotion-best-practices，340 行）+ 35 个细分规则文档（audio/captions/ffmpeg/fonts/lottie/mapbox 等）。在 Remotion monorepo 里维护，没发到 npm，由 Remotion 自家工具链消费。 | You're about to ask the agent to write a non-trivial Remotion component — captions on a clip, a Lottie animation embed, a custom ffmpeg pre-process, or a font-loading setup. You don't want to spend 20 minutes correcting it after, and you don't want to read 35 doc pages yourself.你准备让 agent 写一个非平凡的 Remotion 组件 —— 视频字幕、嵌入 Lottie 动画、自定义 ffmpeg 预处理、字体加载。你不想 agent 写完再花 20 分钟修,也不想自己读 35 页文档。 | Free skill (MIT). Needs your AI agent harness (Claude Code / Cursor / Codex) plus the agent's standard token cost. Package is `private: true` on npm — install means cloning + copying SKILL.md + rules/ into your agent's skills directory; not one command.skill 免费(MIT)。需要 AI agent 平台(Claude Code / Cursor / Codex) + 标准 agent token 成本。包标了 `private: true`(npm 装不到) —— 安装方式是 clone + 把 SKILL.md + rules/ 复制到 agent skills 目录;不是一行命令。 | 69 🛠 Available可使用 | → |
| 19 | zinan92/content-toolkit ↗ | One CLI ('content') that routes URL / video / text / platform-task to the right downstream skill — download, extract, analyze, rewrite, edit, publish, native-XHS — installing each skill on first use.一条 CLI ('content') 把 URL / 视频 / 文本 / 平台任务路由到对应的下游 skill —— 下载、提取、分析、改写、剪辑、发布、小红书原生 —— 每个 skill 第一次用时自动装。 | You see a viral Douyin video. You want to: download it → analyze the hook → rewrite for XHS → cut a 30s clip with subtitles → publish. Right now that's 5 tools with 5 syntaxes. With content-toolkit it's `content download`, `content analyze`, `content rewrite`, `content videocut`, `content publish` — same shape, sub-skills cooperate.看到一个抖音爆款。你要：下载 → 分析 hook → 改写小红书版 → 剪 30s 带字幕版 → 发布。现在这是 5 个工具 5 套语法。用 content-toolkit 是 `content download / analyze / rewrite / videocut / publish` —— 一个形状，下游 skill 协作。 | Free OSS (MIT). Need Node.js 18+, Python 3.11+, ffmpeg, git. ANTHROPIC_API_KEY for analysis + rewrite. First-time install of a skill takes 10-30s (clones repo + venv). Each downstream skill may have its own deps.MIT 免费开源。要 Node.js 18+、Python 3.11+、ffmpeg、git。分析 + 改写要 ANTHROPIC_API_KEY。第一次用某 skill 装 10-30s（clone repo + 建 venv）。下游 skill 可能各自有依赖。 | 67 🛠 Available可使用 | → |
| 20 | oaker-io/wewrite ↗ | A Claude Code / OpenClaw skill that turns 'write a WeChat article' into a full pipeline: scrape live hotspots from Weibo/Toutiao/Baidu, score topics, pick from 7 frameworks, write in 1 of 5 voice personas, SEO-optimize, generate cover + inline images via 9 providers, render with 16 dark-mode-aware themes, and push to WeChat draft box. Around it sits a 6-command CLI for preview, theme gallery, and image post composition.一个 Claude Code / OpenClaw skill，把「写一篇公众号文章」展开成完整管线：从微博 / 头条 / 百度抓实时热搜、做选题打分、从 7 套写作框架挑、用 5 种写作人格中的一种行文、做 SEO 优化、用 9 个 provider 生成封面 + 内文配图、用 16 个支持暗黑模式的主题排版、最后推送到公众号草稿箱。外面包了一个 6 子命令的 CLI 做预览 / 主题画廊 / 图片帖。 | It's morning, a topic is trending on Weibo, you want a 1500-word article with 3 inline images + a cover, in your usual voice, dark-mode-safe formatting, pushed straight to your WeChat draft box where you'll spend 5 minutes adding 2-3 personal sentences before publishing.早上微博上有个热点，你想要一篇 1500 字的文章 + 3 张内文配图 + 封面，按你平常的语气写、暗黑模式不会乱、直接推到公众号草稿箱 —— 然后你花 5 分钟在 2-3 个 anchor 处加上自己的话再发。 | MIT-licensed OSS, free. Runtime cost = your agent's LLM tokens + image-gen provider fees (DashScope/Doubao ~¥0.1/img). WeChat publishing free; needs official-account appid/secret. Without API keys: degrades to local HTML + image prompts (still useful as a draft tool).MIT 协议、开源免费。运行成本 = agent 的 LLM token + 生图 provider 费（DashScope / 豆包大约每张 ¥0.1）。推送公众号免费；要公众号 appid/secret。不配 API key：降级成本地 HTML + 图片提示词（当起稿工具仍然有用）。 | 66 🛠 Available可使用 | → |
| 21 | gooseworks-ai/goose-skills ↗ | An npm-installed catalog of 204 GTM / sales / marketing / competitive-intelligence skills for Claude Code, Cursor, and Codex — broken out into 143 capabilities (atoms), 56 composites (molecules), and 5 playbooks (compounds), each with its own SKILL.md.一个 npm 安装的 GTM / 销售 / 营销 / 竞品分析 skill 目录，目前 204 个 skill：143 个 capability（原子）+ 56 个 composite（分子）+ 5 个 playbook（复合物），适配 Claude Code、Cursor、Codex，每个 skill 有自己的 SKILL.md。 | You want to ask your coding agent "build me a battlecard for Notion vs Linear" or "qualify this inbound lead against my ICP" and have a vetted, reusable skill answer instead of spelling out the workflow from scratch every time.你想跟 coding agent 说「给我做一份 Notion vs Linear 的销售对照卡」或「按我的 ICP 给这条 inbound lead 评分」，让一个经过设计的可复用 skill 来执行，而不是每次从零口述工作流。 | The skill bundle itself is free (npm package declares MIT; repo tree has no LICENSE — caveat). Some composites need third-party API keys (Apollo, Semrush, Ahrefs, Meta Ad Library, Reddit, LinkedIn scrapers via Apify) — those are paid. Per-skill LLM tokens vary; multi-skill playbooks (e.g. event-prospecting-pipeline) are token-heavy.技能包本身免费（npm 包声明 MIT，但仓库树没有 LICENSE 文件 —— 注意）。一部分 composite 要外部 API key（Apollo / Semrush / Ahrefs / Meta Ad Library / Reddit / Apify 上的 LinkedIn 爬虫）—— 那些是收费的。每个 skill 的 LLM token 消耗不同；多 skill 的 playbook（如 event-prospecting-pipeline）token 消耗大。 | 66 🛠 Available可使用 | → |
| 22 | putyy/res-downloader ↗ | A cross-platform desktop downloader (macOS / Linux / Windows) for 9 CN-content sources — 视频号 / 小程序 / 抖音 / 快手 / 小红书 / 直播流 / m3u8 / 酷狗 / QQ音乐. Go + Wails, Apache 2.0, 17K stars.跨平台桌面下载器(Mac / Linux / Win),支持 9 类中文内容源 —— 视频号 / 小程序 / 抖音 / 快手 / 小红书 / 直播流 / m3u8 / 酷狗 / QQ 音乐。Go + Wails,Apache 2.0,17K stars。 | You found a video on Douyin / 视频号 / Bilibili / etc., or you want to archive a 直播 stream, or download a 小红书 image-deck. You want a desktop app where you paste a URL and get the file. No CLI, no browser-extension fiddling, no token-auth dance.你在抖音 / 视频号 / B 站等找到一个视频,或想归档直播流,或下小红书图文。你要一个桌面 app,贴 URL 拿文件。不要 CLI,不要浏览器扩展折腾,不要 token 鉴权。 | Apache 2.0, free. No API keys needed (it consumes public share URLs). Anti-bot risk: pushed 2025-12-31 (~4 months stale); signing for fast-iterating platforms like Douyin may have drifted. Local-only — no cloud.Apache 2.0,免费。不需要 API key(消费的是公开分享 URL)。反爬风险: 2025-12-31 推送(~4 个月没更新);抖音这种快速迭代平台的签名可能已经漂移。纯本地,不上云。 | 65 🛠 Available可使用 | → |
| 23 | zarazhangrui/tab-out ↗ | A Chrome Manifest V3 extension that replaces the new-tab page with a domain-grouped dashboard of all your open tabs — close groups with a swoosh + confetti, save for later, dedup duplicates. Local storage, but every tab's domain leaks to Google's favicon service for icon rendering.一个 Chrome MV3 扩展：把新标签页换成按域名分组的 tab 看板，关掉分组带音效和彩纸，可以收藏 tab，自动去重。数据存本地 chrome.storage.local，但每个 tab 的域名会经 Google 的 favicon 服务渲染图标 —— 实际上 Google 拿得到你打开的所有域名清单。 | It's Friday afternoon. You have 78 tabs open, half are Slack threads, half are GitHub PRs, half are docs you swore you'd read. You hit Cmd+T — instead of an empty new-tab page, Tab Out shows a grid grouped by domain. You triage: close the Slack group (swoosh), save the docs group 'for later', jump to one specific GitHub PR by clicking its tile.周五下午，开了 78 个 tab，一半 Slack 线程，一半 GitHub PR，一半「以后会读」的文档。Cmd+T 一按，Tab Out 给你一个按域名分组的网格 —— 不是空白新标签页。你开始清：Slack 那组关掉（swoosh）、文档那组「保存以后看」、特定的 GitHub PR 点 tile 跳过去。 | Free OSS (MIT licensed). No subscription, no account, no server. Browser must be Chrome / Edge / Brave (anything Chromium-based). Installs as unpacked extension — no Chrome Web Store listing yet.开源免费（MIT 协议）。没订阅、没账号、没服务器。浏览器要 Chrome / Edge / Brave（Chromium 系列都行）。装的是未打包扩展，还没上 Chrome Web Store。 | 65 🛠 Available可使用 | → |
| 24 | tukuaiai/tradecat ↗ | A local-only Python CLI/TUI that pulls 4 public market-data snapshots from Google Sheets, caches them as JSON, and lets you browse them in a terminal panel — no API key, no server, no database, just a single command. Now packaged as a Skill-shell repo with full root governance + CI.一个完全本地的 Python CLI/TUI 工具：从 Google Sheets 拉 4 个公开行情和事件流快照，存为本地 JSON 缓存，在终端面板里浏览。不要 API key，不要服务端，不要数据库，一条命令装完。新版重构成 Skill 外壳 + 项目源的双层结构，根目录带完整治理脚本和 GitHub Actions CI。 | You want to glance at the live event stream and 3 market snapshots from a SSH session or local terminal, with no setup beyond a single install command.你想从 SSH 会话或本地终端瞄一眼实时事件流和 3 个市场快照，除了一条安装命令以外什么都不想配。 | — | 63 🛠 Available可使用 | → |
| 25 | zarazhangrui/frontend-slides ↗ | A Claude Code skill that turns a topic into a polished HTML slide deck — consistent layout, browser-native, deployable to Vercel in one click.一个 Claude Code skill,把一个主题变成一份调性统一的 HTML 演示稿 —— 浏览器原生、可一键部署到 Vercel。 | You have a topic and want a 10-30 page deck out of it. You want it to look polished, run in a browser (so you can host it on Vercel and share a link), and stay editable later by hand without touching a binary file.你有个主题,想做一份 10-30 页的 deck。你希望成品有调性、能在浏览器跑(可以扔到 Vercel 上分享链接)、之后想手改不会卡在 .pptx 二进制里。 | Skill is free. You need a Claude Code subscription (the skill drives the LLM). Optional Vercel free tier if you want one-click deploy.skill 免费。需要 Claude Code 订阅(它驱动 LLM 生成)。可选 Vercel 免费档(想一键部署的话)。 | 62 🛠 Available可使用 | → |
| 26 | karpathy/autoresearch ↗ | Karpathy's autonomous-research template — let your AI agent edit a single 630-line train.py, run 5-minute training experiments overnight, and wake up to a log of 100 attempts and (hopefully) a better model. The 'program.md' file is the lightweight skill that drives the agent.Karpathy 出的「让 AI agent 替你做研究」模板 —— 让 agent 自己改一个 630 行的 train.py，每次跑 5 分钟训练实验，过夜跑 100 轮，早上起来看日志和（希望中的）更好的模型。program.md 是驱动 agent 的轻量 skill。 | It's 10pm. You spin up Claude Code or Codex in the autoresearch repo, point it at program.md, set a token + GPU-time budget, and go to sleep. By morning the agent has run ~100 5-minute training experiments, kept the architectural / hyperparameter changes that lowered val_bpb, and left you a journal you can read with coffee.晚上 10 点，你在 autoresearch 仓库里启动 Claude Code 或 Codex，指向 program.md，设好 token + GPU 时间预算，睡觉。早上起来，agent 已经跑了大约 100 次 5 分钟训练实验，留下了让 val_bpb 下降的架构 / 超参数变更，并附上你可以喝咖啡读的日志。 | Code is "MIT" per README (no LICENSE file shipped — caveat). Hardware: single NVIDIA H100 — cloud H100 ≈ $2-4/hr depending on provider, ≈ $20-40 for an overnight run. Agent tokens: each experiment iteration burns coding-agent tokens; budget for 100 experiments before launch.代码按 README 说是 "MIT"（仓库没有 LICENSE 文件 —— 注意）。硬件：单卡 NVIDIA H100 —— 云 H100 大概 $2-4/小时，通宵跑大约 $20-40。Agent token：每次实验都烧 coding-agent token；100 个实验的预算先估好再启动。 | 61 🛠 Available可使用 | → |
| 27 | zarazhangrui/codebase-to-course ↗ | A Claude Code skill that turns any codebase into a single-page interactive HTML course — scroll-based modules, animated data-flow diagrams, code↔plain-English side-by-side, embedded quizzes, glossary tooltips. Skill ships real CSS / JS / build.sh, not just a prompt; output is a directory you open in any browser.一个 Claude Code 技能：把任意代码库变成一个单页交互式 HTML 课程 —— 滚动式模块、数据流动画、代码 ↔ 大白话对照、嵌入式 quiz、术语 tooltip。skill 自带真的 CSS / JS / build.sh，不只是 prompt；产物是一个目录，浏览器打开就能用。 | You forked an open-source repo (or your AI agent built one for you) and now you need to actually understand it before adding features, debugging it, or explaining it to a teammate. You don't have time for a CS course — you need a 30-minute scroll-through that traces what happens when the app actually runs.你 fork 了一个开源仓库（或者 agent 给你生了一个），现在要加功能 / debug / 跟同事讲解之前必须真的懂它。没时间上 CS 课，要的是 30 分钟的滚动式讲解，跟着 app 真实运行的路径走一遍。 | Skill itself is free OSS (no LICENSE though). You need a Claude Code subscription. Per-course token cost depends on codebase size — typically a few dollars for a medium repo.skill 开源免费（虽然没 LICENSE）。需要 Claude Code 订阅。每生成一份课程的 token 成本看代码库大小 —— 中等仓库通常几美元。 | 60 🛠 Available可使用 | → |
| 28 | zarazhangrui/follow-builders ↗ | A Claude Code / OpenClaw skill that delivers a daily or weekly digest of what 25 top AI builders are saying — summaries of new podcast episodes (Latent Space, No Priors, Training Data, etc.), insights from X posts, and Anthropic / Claude blog articles. Bilingual EN / ZH. No API keys needed (centrally fetched feeds).一个 Claude Code / OpenClaw 技能：定期送上 25 位顶级 AI builder 的内容摘要 —— 顶尖 AI 播客（Latent Space / No Priors / Training Data 等）+ X 上的精华帖 + Anthropic / Claude 官方博客。中英双语，不需要 API key（feed 由项目方中心化抓取）。 | Monday morning, you want a single message summarizing what the 25 top AI builders said over the weekend — what Karpathy posted, what was new on Latent Space, what Anthropic Engineering shipped — bilingual if you read both EN/ZH. Delivered to Telegram or in-chat, no logging into 6 platforms.周一早上想一条消息看完 25 位顶尖 AI builder 周末说了什么 —— Karpathy 发了啥、Latent Space 出了什么新一期、Anthropic Engineering 写了什么 —— 中英双语都看。送到 Telegram 或直接在 chat 里，不用登 6 个平台。 | Free OSS (no LICENSE though). Need a Claude Code or OpenClaw subscription. No third-party API keys required for content (Zara's pipeline pays that cost), but you do need the agent host to run. Optional Telegram / Discord / email for delivery (free tiers fine).开源免费（虽然没 LICENSE）。需要 Claude Code 或 OpenClaw 订阅。内容不要第三方 API key（Zara 那边掏钱），但 agent host 得跑。可选 Telegram / Discord / 邮箱投递（免费档够用）。 | 60 🛠 Available可使用 | → |
| 29 | RKiding/Awesome-finance-skills ↗ | A 10-skill bundle for Antigravity / OpenCode / OpenClaw that turns your AI agent into a finance analyst — fetches multi-source news (Cailian, Polymarket, etc), A/HK/US stock data, runs FinBERT sentiment, draws transmission-chain diagrams (draw.io XML), and writes structured reports. Each skill is SKILL.md + Python scripts.一组 10 个金融分析 skill，适配 Antigravity / OpenCode / OpenClaw —— 把 AI agent 变成金融分析员：拉多源新闻（财联社、Polymarket 等）、A 股 / 港股 / 美股行情、FinBERT 情绪打分、画市场传导链图（draw.io XML）、写结构化报告。每个 skill 是 SKILL.md + Python 脚本的组合。 | An overnight event (Polymarket odds shift / a Cailian breaking news flash / a US session crash) hits, you want your agent to: (a) pull the news, (b) score sentiment, (c) draw a transmission chain to A-shares / HK / US, and (d) hand back a draftable report — without you scaffolding 4 separate prompts.夜里出了个事（Polymarket 赔率变了 / 财联社快讯 / 美股盘中跳水），你想让 agent：(a) 拉新闻、(b) 打情绪分、(c) 画传导链到 A 股 / 港股 / 美股、(d) 给出可改的研报草稿 —— 不想自己分 4 段 prompt 搭场子。 | Apache 2.0, free OSS. Runtime cost = your agent's normal LLM token usage + per-skill external service fees: news scraping is mostly free (Cailian / Polymarket public endpoints), FinBERT runs locally on CPU (free), Kronos forecasting is local (free). alphaear-search may use Jina / DDG / Baidu — Jina has paid tier, DDG/Baidu free.Apache 2.0、开源免费。运行成本 = agent 正常 LLM 用量 + 每个 skill 自带的外部服务费：新闻抓取大多免费（财联社 / Polymarket 公开端点）、FinBERT 本地 CPU 跑（免费）、Kronos 预测本地跑（免费）。alphaear-search 可能用 Jina / DDG / 百度 —— Jina 有付费档，DDG / 百度免费。 | 59 🛠 Available可使用 | → |
| 30 | autoclaw-cc/xiaohongshu-skills ↗ | A 5-skill OpenClaw / Claude Code bundle that automates Xiaohongshu (小红书) — login, publish, search, comment/like/bookmark, and orchestrated workflows like trend-tracking — by driving the user's real Chrome session through a paired browser extension. No XHS API key, no encryption reversal.一组 5 个 OpenClaw / Claude Code 技能：在小红书做登录、发布、搜索、评论 / 点赞 / 收藏 + 复合运营（追热点、竞品、批量互动）。通过自带的 Chrome 扩展驱动你自己的真实小红书账号，不需要 XHS API key，也不用逆向加密。 | You want to tell your agent "find the top assassin's-creed image notes, bookmark them and tell me what they're about" and have it actually search XHS, filter to image notes, sort by likes, bookmark, fetch detail, and summarize — without you opening 20 tabs by hand.你想跟 agent 说「找刺客信条最火的图文，收藏并告诉我讲了什么」，然后它真的会去搜小红书、筛图文、按点赞排序、收藏、拉详情、做总结 —— 不用你自己开 20 个 tab 手动操作。 | Free OSS (MIT). Need Python 3.11+, uv, Chrome with the unpacked XHS Bridge extension loaded, and a real XHS account. Agent-side LLM cost is on you (Anthropic Claude / OpenClaw provider).开源免费（MIT）。需要 Python 3.11+、uv、加载了 XHS Bridge 解压扩展的 Chrome、一个真实 XHS 账号。Agent 端 LLM 成本自付（Anthropic Claude / OpenClaw 提供商）。 | 59 🛠 Available可使用 | → |
| 31 | zinan92/doc-driven-dev-workflow ↗ | A terminal-first, repo-local, doc-driven development workflow — 5 phases × 22 stages canonicalised in JSON, with task scaffolding scripts, a workflow_guard state machine that rejects invalid moves, and a React observer dashboard the human watches while Codex + Claude Code execute against the spec.一个终端优先、仓库本地、文档驱动的开发工作流 —— 5 阶段 × 22 stage 的 canonical JSON,task 脚手架脚本,工作流闸门(workflow_guard)拒绝非法状态迁移,加一个 React observer dashboard 让你边看边让 Codex + Claude Code 按 spec 干活。 | You're about to give Codex + Claude Code a non-trivial task — a new feature, a refactor, a migration. You know the agents will jump to coding without a research / PRD step. You want a structured handoff: research outputs, then PRD, then plan, then bounded code batches with reviews, with approval gates between phases that you (the human) hold.你准备给 Codex + Claude Code 一个不平凡的任务 —— 新功能 / 重构 / 迁移。你知道 agent 会直接跳到写代码,跳过研究 / PRD 步骤。你想要结构化的 handoff: 研究产出 → PRD → 实现计划 → 限定批次的代码 + 评审,phase 之间有你(人)把控的审批门。 | No LICENSE file yet (treat as personal-use until upstream ships one). Free OSS otherwise. Need Python 3.9+ + Node 18+ (frontend dashboard). You bring your own AI agent subscriptions (Codex + Claude Code or equivalents). No external services at runtime — everything is repo-local files + a localhost dashboard.暂无 LICENSE 文件(在上游补之前当个人用看待)。其他都免费开源。需要 Python 3.9+ + Node 18+(前端 dashboard)。自带 AI agent 订阅(Codex + Claude Code 或同类)。运行时不需要外部服务 —— 都是仓库本地文件 + 本机 dashboard。 | 59 🛠 Available可使用 | → |
| 32 | Jamailar/RedBox ↗ | An Electron desktop AI workspace for Xiaohongshu creators — built-in browser captures content from XHS / YouTube / web into a local knowledge base, AI editor drafts manuscripts, RedClaw automation console runs scheduled / long-running tasks, and image+video generation feeds the publish pipeline. BYO LLM endpoint + key.一个面向小红书创作者的 Electron 本地 AI 工作台 —— 内置浏览器从小红书 / YouTube / 网页抓内容进本地知识库，AI 编辑器写稿，RedClaw 自动化控制台跑定时和长周期任务，生图 + 生视频接发布链路。LLM endpoint + key 自带。 | Tuesday morning you save 8 trending XHS notes via the Plugin, "wander" them in the local KB to spark a topic, draft 3 manuscripts in the editor, generate cover images, queue them in RedClaw to publish across the week, then close the laptop — RedClaw keeps running.周二早晨用插件保存 8 条热门小红书笔记进本地知识库，「漫步」碰撞出选题，在编辑器里起 3 篇稿、生封面图、扔进 RedClaw 队列让它一周内分批发出，关掉笔记本 —— RedClaw 在后台继续跑。 | App is free for personal use (LICENSE is MIT-NC, no commercial use). You bring your own LLM endpoint + key — Anthropic / OpenAI / Google / openai-compatible all supported via Vercel AI SDK v6. Image / video generation costs depend on the model you point it at (e.g. GPT-image-2 is paid).App 个人用免费（LICENSE 是 MIT-NC，禁商用）。LLM endpoint + key 自带 —— 通过 Vercel AI SDK v6 支持 Anthropic / OpenAI / Google / openai-compatible。生图 / 生视频成本看你接哪家模型（比如 GPT-image-2 是付费的）。 | 58 🛠 Available可使用 | → |
| 33 | nicobailon/visual-explainer ↗ | An agent skill that turns 'draw a diagram of X' into a self-contained styled HTML page (with Mermaid + Chart.js + dark/light themes) that opens in your browser, instead of an ASCII-art mess in the terminal. 8 slash commands, 4 reference templates, multi-harness support (Claude Code / Pi / Codex / OpenCode / Cursor / OpenClaw).一个 agent skill —— 把「画一张 X 的图」从终端里那堆 ASCII 字符变成自包含的样式化 HTML 页面（含 Mermaid + Chart.js + 明暗主题），自动在浏览器里打开。8 个 slash command、4 个参考模板，跨 Claude Code / Pi / Codex / OpenCode / Cursor / OpenClaw 6 个 host 都能装。 | You ask your agent: 'compare these 4 architecture options across 6 dimensions' or 'review this 200-line refactor diff'. Without this skill you get a wrapped, broken ASCII table or a mermaid block that won't render in your terminal. With it, an HTML page opens in your browser with proper typography + Mermaid SVG + dark/light theme.你跟 agent 说：「比一下这 4 个架构方案在 6 个维度上的差异」或「review 一下这 200 行 refactor 的 diff」。没装这个 skill：终端里一堆换行错乱的 ASCII 表 / 不能渲染的 mermaid block。装了：浏览器里直接开一个 HTML 页面，排版整齐 + Mermaid SVG + 明暗主题。 | MIT-licensed, free OSS. Generation cost is your agent's normal LLM cost (a single command can spend several thousand tokens reading templates + producing HTML). /share-page additionally requires the separate `vercel-deploy` skill + a Vercel account.MIT 协议、开源免费。生成成本就是你 agent 正常的 LLM 用量（单条命令读模板 + 出 HTML 大概几千 token）。/share-page 额外要装 `vercel-deploy` skill + Vercel 账号。 | 57 🛠 Available可使用 | → |
| 34 | zarazhangrui/personalized-podcast ↗ | A Claude Code skill (`/podcast`) that turns any content (URL / file / text / a topic) into a 2-host AI conversation episode — Claude writes the script, Fish Audio voices it, pydub stitches the audio, optional RSS so it shows up in Apple Podcasts / Spotify / Overcast next to your real subscriptions.一个 Claude Code 技能（`/podcast`）：把任何内容（URL / 文件 / 文字 / 主题）变成两位 AI 主持人对谈的播客 episode —— Claude 写台词，Fish Audio 配音，pydub 拼音频，可选 RSS 让它和你订阅的播客一起出现在 Apple Podcasts / Spotify / Overcast 里。 | A 4000-word newsletter just landed in your inbox. You'd rather listen on tomorrow's commute than read it tonight. You paste the URL into `/podcast`, by morning there's an MP3 in your podcast app — two hosts walking through it. Or weirder: you paste your own resume and ask the hosts to comment on your career.一封 4000 字的 newsletter 刚到 inbox，你想明天通勤路上听，不想今晚硬读。把 URL 丢给 `/podcast`，第二天早上播客 app 里就有一份 MP3 —— 两位主持人帮你走一遍。或者更怪一点：把自己简历丢进去，让主持人点评你的职业轨迹。 | Skill is free OSS (no LICENSE though). Need Claude Code subscription (~$0.10 / 10-min episode in Claude tokens). Fish Audio free tier sufficient for personal use. Optional GitHub Pages (free) to host the RSS feed.skill 开源免费（没 LICENSE）。需要 Claude Code 订阅（10 分钟一集大约 $0.10 token 成本）。Fish Audio 免费档够个人用。可选 GitHub Pages（免费）托管 RSS feed。 | 54 🛠 Available可使用 | → |
| 35 | zarazhangrui/youtube-to-ebook ↗ | A Python tool + Claude skill that pulls latest videos from YouTube channels you follow, runs the transcripts through Claude to make magazine-style articles, packages them as an EPUB delivered to your email or Apple Books — set it on a Mac LaunchAgent for weekly auto-delivery.一个 Python 工具 + Claude 技能：从你订阅的 YouTube 频道拉最新视频，把字幕用 Claude 改写成杂志体文章，打成 EPUB 邮件给你或导入 Apple Books。配 Mac LaunchAgent 就能每周自动送到。 | Sunday evening. A weekly EPUB lands in your email — last week's episodes from your 5 favorite YouTubers, rewritten as readable magazine articles by Claude. Apple Books picks it up automatically. Monday commute = quiet reading; no autoplay, no related-videos spiral, no algorithm stealing 90 minutes.周日晚上，邮箱收到一份 EPUB —— 关注的 5 个 YouTuber 这周的视频，Claude 改写成杂志体可读文章。Apple Books 自动接收。周一通勤 = 安静读完；没有自动播放、没有相关视频推荐、没有算法偷你 90 分钟。 | Free OSS (no LICENSE though). YouTube Data API free tier (10K units/day) is plenty for personal use. Anthropic Claude is BYOK — typically a few dollars per weekly book depending on episode count + transcript length. Email delivery via your existing SMTP (Gmail / iCloud).开源免费（没 LICENSE）。YouTube Data API 免费档（每日 10K 配额）个人用绰绰有余。Anthropic Claude 自带 key —— 一般每周书几美元，看 episode 数和字幕长度。邮件投递用现有 SMTP（Gmail / iCloud）。 | 53 🛠 Available可使用 | → |
| 36 | zinan92/content-downloader ↗ | Single command, 4 Chinese-content platforms (Douyin / XHS / WeChat / X), one normalized output shape — but at eval time WeChat + X silently produce empty files while reporting success.一条命令搞定 4 个中文内容平台（抖音 / 小红书 / 公众号 / X），输出格式统一 —— 但测评时公众号和 X 在「成功」表象下产出空文件。 | You build a content pipeline that needs to ingest video / posts / articles across multiple Chinese platforms. You want one CLI shape, one output schema, one manifest — not 4 platform-specific scripts.你在搭一条内容管线，要从多个中文平台采素材（视频 / 帖子 / 文章）。你要的是一份 CLI 形状、一份输出 schema、一份 manifest —— 不是 4 套各自一套的脚本。 | Free OSS (MIT). Need Python 3.11+, Playwright (auto-installs Chromium), yt-dlp (auto-installs). For Douyin: prepare cookies.json. For XHS: sidecar auto-installs. WeChat + X currently produce broken output.MIT 免费开源。要 Python 3.11+、Playwright（自动装 Chromium）、yt-dlp（自动装）。抖音：准备 cookies.json。小红书：sidecar 自动装。公众号 + X 当前产出有问题。 | 44 ⚠️ Risky有风险 | → |
| 37 | zinan92/videocut ↗ | A 7-step CLI pipeline for AI-driven 口播 (talking-head) video editing — cuts filler words, generates SRT captions, extracts quote clips + chapter splits + cover cards, applies speed adjustment. Each step usable standalone; pipeline.js chains them.AI 口播视频编辑 7 步 CLI 流水线 —— 去废话、出字幕、提金句、拆条、做封面、变速。每一步可单独用,pipeline.js 串联起来跑。 | You finished recording a 5-15 minute talking-head video. Now you need: cleaned cut without filler words, SRT subtitles, 3-5 quote clips for TikTok / Douyin / Xiaohongshu cross-post, cover cards for thumbnails, and a speed-adjusted version for "fast watch" channels. Currently this is a multi-tool 2-hour grind.你刚录完一个 5-15 分钟的口播视频。现在要: 去废话的剪辑版、SRT 字幕、3-5 个金句剪辑(给抖音 / 小红书复用)、封面卡片、变速版给"倍速观看"频道。当前是多工具 2 小时手工。 | No LICENSE file (treat as personal-use until upstream commits one). Free OSS otherwise. Need Node 18+ + FFmpeg + Whisper installed locally + Claude CLI subscription (drives the AI decisions). All processing local — no cloud uploads.无 LICENSE 文件(在上游补之前当个人用看待)。其他免费开源。需要 Node 18+ + FFmpeg + Whisper 本地安装 + Claude CLI 订阅(驱动 AI 决策)。处理全部本地,不上云。 | 43 ⚠️ Risky有风险 | → |
| 38 | dreammis/social-auto-upload ↗ | A Playwright-based CLI that auto-uploads videos to 7 Chinese + international social platforms (抖音 / Bilibili / 小红书 / 快手 / 视频号 / 百家号 / TikTok) — login via cookie / QR code, schedule for later, drive from one config. 10.6K stars, MIT-implied (no LICENSE file), single-author / single-commit visible history.一个基于 Playwright 的 CLI,把视频自动上传到 7 个中外社交平台(抖音 / B 站 / 小红书 / 快手 / 视频号 / 百家号 / TikTok) —— cookie / 扫码登录,可定时发布,一份配置驱动多平台。10.6K stars,MIT 暗示但实际无 LICENSE 文件,单作者 / git 仅 1 个 commit 可见。 | You finished a video. You want to push it to 抖音 / Bilibili / 小红书 / 快手 / 视频号 / 百家号 / TikTok with a single command, ideally with platform-specific tweaks (different titles, thumbnails, schedule times). You want to do this every week, repeatedly, without re-learning each platform's UI.视频做完了。你想一行命令推到 抖音 / B 站 / 小红书 / 快手 / 视频号 / 百家号 / TikTok,理想情况下每个平台还能微调(不同标题、缩略图、定时)。你想每周重复一次,不用每次重新摸每家平台的 UI。 | Free OSS but unlicensed (no LICENSE file at repo root despite 10K+ stars). Need Python 3.10+ + Playwright. You bring your own real account on each target platform (cookie / QR login). No LLM cost — purely browser automation. Anti-bot risk: aggressive scheduling can get an account limited; the included stealth.min.js was last regenerated 2024-06-10 (~16 months stale).免费开源但 license 缺失(10K+ stars 但无 LICENSE 文件)。需要 Python 3.10+ + Playwright。每个目标平台带自己的真实账号(cookie / 扫码)。不需要 LLM —— 纯浏览器自动化。反爬风险: 频次过高账号会被限;自带的 stealth.min.js 上次更新 2024-06-10(~16 个月没更新)。 | 42 ⚠️ Risky有风险 | → |
| 39 | zinan92/content-extractor ↗ | Drop in a content-downloader output dir (or a bare .mp4) and get back transcript.json + analysis.json + structured_text.md. Video path works well; audio adapter is documented but unimplemented.丢进 content-downloader 输出目录（或裸 .mp4 文件），拿回 transcript.json + analysis.json + structured_text.md。视频路径稳，但音频 adapter README 说了没实现。 | You downloaded 50 trending videos with content-downloader. Now you want to know which ones are worth watching, what topics they cover, what hooks they use — without scrubbing through 5 hours of video.你刚用 content-downloader 抓了 50 个热门视频。现在想知道哪些值得看、它们讲了什么主题、用了什么 hook —— 不想真的拖进度条看 5 小时。 | Free OSS (MIT). Need Python 3.13+, ffmpeg, ANTHROPIC_API_KEY for analysis (optional — extraction without LLM still produces transcript). Whisper runs locally — no per-minute fees. On Mac, expect real-time transcription (no GPU acceleration).MIT 免费开源。要 Python 3.13+、ffmpeg、ANTHROPIC_API_KEY 用于 LLM 分析（可选 —— 不调 LLM 也能出转录）。Whisper 本机跑，没有按分钟计费。Mac 上实时速度（没有 GPU 加速）。 | 42 ⚠️ Risky有风险 | → |
| 40 | zinan92/kline ↗ | A multi-asset OHLCV candle service for the Park-trading pipeline — one ticker + one timeframe, returns standardised candles. 4 providers: 𝘁𝘂𝘀𝗵𝗮𝗿𝗲 (A股) / 𝘆𝗮𝗵𝗼𝗼 (美股) / 𝗯𝗶𝗻𝗮𝗻𝗰𝗲 (crypto) / 𝗰𝗼𝗺𝗺𝗼𝗱𝗶𝘁𝘆.Park 交易管线的多资产 OHLCV 蜡烛数据服务 —— 一个 ticker + 一个 timeframe,返回标准化蜡烛。4 个 provider: 𝘁𝘂𝘀𝗵𝗮𝗿𝗲(A 股) / 𝘆𝗮𝗵𝗼𝗼(美股) / 𝗯𝗶𝗻𝗮𝗻𝗰𝗲(加密) / 𝗰𝗼𝗺𝗺𝗼𝗱𝗶𝘁𝘆(商品)。 | Your trading pipeline needs daily candles for a watchlist that mixes 中国 A股 + 美股 + crypto + 商品. Each vendor has its own SDK + auth + rate-limit + symbol convention. You want a single HTTP call shape that abstracts over all of it.你的交易 pipeline 要每日蜡烛 —— watchlist 混合 A 股 + 美股 + 加密 + 商品。每家 vendor 有自己 SDK / 鉴权 / rate limit / 符号规范。你要一个统一的 HTTP 调用形状,把这些抽象掉。 | No LICENSE file. Free OSS otherwise. Need Python 3.11+ + the relevant vendor credentials per asset class — TuShare token for A股 is paid; yahoo / binance / commodity have free public access. Self-hosted; no cloud uploads.没有 LICENSE 文件。其他免费开源。需要 Python 3.11+ + 各资产类对应 vendor 凭证 —— A 股的 TuShare token 要付费;yahoo / binance / 商品 公开免费可访问。自托管,不上云。 | 42 ⚠️ Risky有风险 | → |
| 41 | zinan92/AI-videos ↗ | A Python pipeline that turns a character PNG + outfit PNG + motion-reference MP4 into an AI-generated video — drives RunningHub v2 API, logs the run to JSON.一个 Python pipeline,把人物 PNG + 服装 PNG + 动作参考 MP4 变成 AI 生成的视频 —— 驱动 RunningHub v2 API,运行日志写 JSON。 | You have a character image, an outfit you want them to wear, and a motion-reference video showing the kind of action you want. You want the finished video — with that character in that outfit doing that action — without API plumbing.你有一个角色图、想给它穿的服装图、和一个表示要做什么动作的参考视频。你想直接拿到成品 —— 角色穿着这套衣服做那个动作的视频 —— 不要管 API 接线。 | No LICENSE file (treat as personal-use until upstream commits one). RunningHub API needs credits — pay-per-generation; budget USD 1-5 per finished short video depending on quality settings. Local: Python 3.9+.无 LICENSE 文件(在上游补之前当个人用看待)。RunningHub API 要 credits —— 按生成计费;一条成品短视频预算 1-5 美元,看质量设置。本地: Python 3.9+。 | 38 ⚠️ Risky有风险 | → |