
Great AI tools,
right in your hands.把好用的 AI,
直接放到你手里。
News, tutorials, extensions, tools and skill packs — six sections refreshed every day.
See what changes, and take home something you can use right now.
资讯 · 教程 · 插件 · 工具 · 技能包,六大板块每天上新。
看见变化,也拿到能立刻上手的东西。
AI 快讯AI News · Updated daily每日自动更新 · 2026年10月12日


Apple releases tvOS 27.2.1 update, adding Siri AI support to second- and third-generation Apple TV 4K苹果推送 tvOS 27.2.1 更新,第二 / 第三代 Apple TV 4K 新增 Siri AI 支持

Achieve "laid-back office work": Perplexity for Mac launches "AI Teammate" to build your round-the-clock digital workforce.实现“躺平式办公”:Mac 版 Perplexity 上线“AI 队友”,组建你的全天候数字员工

DeepSeek's front-end code shows a "quota exhausted" prompt, leading developers to speculate that a consumer-facing paid plan will be introduced for the first time.DeepSeek 前端代码出现“额度耗尽”提示,开发者猜测将首次推出 C 端付费方案

Eli Lilly CEO responds to "AI will conquer cancer within a decade": The vision is great, but it's unlikely礼来 CEO 回应“AI 十年内攻克癌症”:愿景很美好,但不太可能

Elon Musk shows support for Jensen Huang: He will sign a petition to support his return to the position of CEO of NVIDIA.马斯克声援黄仁勋:将签署请愿,支持其重返英伟达 CEO 岗位

Epic update to local inference engine llama.cpp 1.0.0: Supports over 150 model architectures, and large models can run at full capacity even when offline.本地推理引擎 llama.cpp 1.0.0 史诗级更新:支持超 150 种模型架构,断网也能满血跑大模型

To use general-purpose AI like GPT-6 Astra as the "brain" of robots: OpenAI announces the reorganisation of its embodied intelligence team要让 GPT-6 Astra 等通用 AI 当机器人“大脑”:OpenAI 宣布重组具身智能团队
今日正能量每天进步一点点,AI 陪你把想法变成现实。从一篇图文教程开始 →广告AI 教程AI Tutorials · Step-by-step guides you can follow手把手图文教程,跟着做就会

Illustrated Tutorial: Install claude-obsidian in three steps, so that every AI note has a source and every sentence can be traced back.图文教程:三步装好 claude-obsidian,让 AI 笔记条条有出处、句句可回溯
Follow along: Prepare the Obsidian vault and Claude Code → Install the plugin and connect the skills → Drop a piece of material into the vault and see it generate notes with citations → Ask questions to verify evidence-based answers. Runs entirely locally, making it suitable for users who want to seriously build a second brain.
跟着做一遍:准备 Obsidian 库和 Claude Code → 安装插件并接入技能 → 丢一篇资料进库看它生成带引用的笔记 → 提问验证证据式回答。全程本地运行,适合想认真搭第二大脑的用户。

Illustrated Tutorial: Create Your Product's First Code-Level Promotional Video with video-shotcraft – From Setting Up the Environment to the Final Output图文教程:用 video-shotcraft 给你的产品做第一条代码级宣传片,从装环境到出片全流程
Follow along: Install the skills and Remotion dependencies → Prepare real page assets → Have the Agent storyboard the script → Render the final video and export the CapCut draft for fine-tuning. Take a look at the official Gallery before you start, and you'll be twice as efficient.
跟着做一遍:装好技能与 Remotion 依赖 → 准备真实页面素材 → 让 Agent 分镜出脚本 → 渲染成片并导出剪映草稿精修。先看一遍官方 Gallery 再动手,效率翻倍。

Illustrated Tutorial: Install Microsoft's official TokenEconomizer in three minutes to let your coding agent automatically save money图文教程:三分钟装上微软官方 TokenEconomizer,让编码 Agent 自动省钱
Follow these steps: Extract the plugin source code package → Register with Claude Code using three commands → Open the import for historical sessions → Understand the status report → Adjust parameters as needed. Requires Python 3.10+. Supports Claude Code, Codex, GitHub Copilot CLI, and Pi. No API key required at any point.
跟着做一遍:解压插件源码包 → 三条命令注册进 Claude Code → 打开历史会话导入 → 看懂 status 报告 → 按需调参。需要 Python 3.10+,支持 Claude Code、Codex、GitHub Copilot CLI 与 Pi,全程无需任何 API Key。

Illustrated Tutorial: A Step-by-Step Guide to Conducting Your First Six-Stage Security Audit Using Official Cloudflare Techniques图文教程:手把手教你用 Cloudflare 官方技能跑第一次六阶段安全审计
Follow along step by step: Unzip and install the security-audit skill → Select a small repository for a test run → Understand the division of labour between the reconnaissance ledger and vulnerability hunting → Comprehend the three types of determinations in findings.json and the final report → Master incremental auditing and cost control. Compatible with Claude Code / Codex / Cursor, with zero additional dependencies.
跟着做一遍:解压安装 security-audit 技能 → 选一个小仓库试跑 → 看懂侦察账本与猎洞分工 → 读懂 findings.json 三种判定与最终报告 → 掌握增量审计与成本控制。适配 Claude Code / Codex / Cursor,零额外依赖。
AI 插件AI Extensions · Browser extensions with official packs浏览器插件,附官方安装包

Claude-Obsidian: Turn Claude Code and Obsidian into a "local-first" knowledge work partnership. Any information you put in becomes a note with a reference link.claude-obsidian:让 Claude Code 和 Obsidian 变成一对「本地优先」知识工作搭档,任何资料丢进去都变成带引用链接的笔记
Do you have two hundred unread articles piled up in your favourites and dozens of PDFs sitting in your downloads folder, and you want AI to help you process them but are always worried that it will make up sources? This open-source project with over 16,400 stars directly connects Claude Code (or any host compatible with Agent Skills) to the Obsidian vault: you feed it any source, and it generates interlinked Markdown notes with each conclusion citing its source. When you ask questions, the answers are strictly based on the evidence already in the vault, and if it can't find an answer, it simply says so. It runs entirely locally, and your data never leaves your computer.
收藏夹里堆了两百篇没读的文章、下载目录里躺着几十份 PDF,想让 AI 帮你消化却总担心它瞎编出处?这个 16,400+ 星的开源项目把 Claude Code(或任何兼容 Agent Skills 的宿主)和 Obsidian 库直接接通:你把任何来源丢进去,它产出互相链接、每个结论都标明出处的 Markdown 笔记;提问时回答也严格基于库里已有证据,不给答案就直说找不到。全程本地运行,数据不出你的电脑。

Microsoft's Official Open-Source TokenEconomizer: Without Changing a Single Line of Code in the Agent, Actual Coding Costs Are Reduced by 30.9%微软官方开源 TokenEconomizer:不改 Agent 一行代码,实测编码成本直降 30.9%
Presented by UC Berkeley in collaboration with Microsoft Research: Learn three techniques from your past sessions – micro-skills to remember repository rules, eliminate tools that have never been used, and stop failure loops that go round and round. SWE-Bench tested 277 real-world tasks, and all 30 "repository × model" combinations were cheaper, with a slight increase in task success rate. Supports Claude Code, Codex, GitHub Copilot CLI, and Pi.
UC Berkeley 联手微软研究院出品:从你跑过的历史会话里学三招——微技能记住仓库规矩、裁掉从没用的工具、拦住原地打转的失败循环。SWE-Bench 277 个真实任务实测,30 个「仓库×模型」组合全部更便宜,任务成功率还略有提升。支持 Claude Code、Codex、GitHub Copilot CLI、Pi。
stop-that-shit: Chinese open source "Mission Boundary Guard", a red stop seal to shut up when the AI is donestop-that-shit:中文开源「任务边界守卫」,一个红色 STOP 印章让 AI 干完正事就闭嘴
You only ask the Agent to export one file, and it calculates the SHA-256 that no one reads, saying "just in case"; you let it do only Review, and it finds the problem and starts to change the code. This Chinese open source plugin, whose name is even straightforward enough, uses a set of Skill plus Guard hooks to stop the four types of redundant actions of AI in real time in mainstream hosts such as Codex, Claude Code, OpenCode, and Hermes Agent CLI - range expansion, useless defense, intention to cross the border, and task rotation. The official benchmark can save about 65% of the nonsense output tokens, and the technical content is not discounted at all.
你只让 Agent 导出一个文件,它顺手算了份没人读取的 SHA-256,说是「以防万一」;你让它只做 Review,它找到问题就开始改代码。这个连名字都够直白的中文开源插件,用一套 Skill 加 Guard 钩子,在 Codex、Claude Code、OpenCode、Hermes Agent CLI 等主流宿主里实时拦下 AI 的四类多余动作——范围膨胀、无用防御、意图越界、任务打转,官方基准能省约 65% 的废话输出 token,而且技术内容一点不打折。

watermarks-remover: 25,000-star open source plugin to clean up the AI watermarks Claude, Gemini, OpenAI buried in your fileswatermarks-remover:2.5 万星开源插件,把 Claude、Gemini、OpenAI 埋在你文件里的 AI 水印一次性剥干净
The documents you ask AI to help write and the images you generate may contain invisible "ID cards": invisible Unicode characters, statistical text watermarks, C2PA/EXIF metadata, all quietly marked "This content is generated by AI". This open source project in October studied the traceability marks of major manufacturers thoroughly, added a proxy skill to the service implemented by a pure Python standard library, and cleaned them on your own computer - no online uploads, no touching other people's content, and focused on "only handling your own things".
你让 AI 帮忙写的文档、生成的图片,里面可能藏着看不见的「身份证」:不可见的 Unicode 字符、统计型文本水印、C2PA/EXIF 元数据,都在悄悄标注「这段内容由 AI 生成」。这个 10 月蹿红的开源项目把各大厂商的溯源标记研究了个透,用一个纯 Python 标准库实现的服务加一个代理技能,在你自己的电脑上把它们清干净——不联网上传、不碰别人的内容,主打一个「只处理你自己的东西」。
AI 工具AI Tools · Desktop & CLI tools, ready to download桌面与命令行工具,直接下载

browser-use: The browser automation champion with 119,000 stars releases its stable version 1.0. AI-controlled browsers finally have a de facto standard.browser-use:11.9 万星的浏览器自动化王者发布 1.0 稳定版,AI 操控浏览器这件事终于有了事实标准
Enabling an AI Agent to open a browser, click buttons, fill out forms, and scrape data is something that browser-use has achieved as the most widely cited open-source implementation in the entire field – with 119,000 stars on GitHub. Over the past year, almost all "AI-controlled browser" projects have been built on top of it. The project has now released the 1.0.4 stable version: rewritten based on Pydantic AI, with a persistent Python REPL running directly on the native CDP protocol. It can take over your local Chrome and also supports cloud browsers, with both speed and stability improved.
让 AI Agent 自己开浏览器、点按钮、填表单、抓数据,这件事 browser-use 已经做成了整个领域引用最多的开源实现——GitHub 11.9 万星,一年来几乎所有「AI 操控浏览器」项目都在它之上搭建。现在项目发布 1.0.4 稳定版:基于 Pydantic AI 重写,一条持久的 Python REPL 直接跑在原生 CDP 协议上,可接管你本地的 Chrome,也支持云端浏览器,速度和稳定性都上了一个台阶。

Vercel agent-desktop: A command-line tool that enables AI agents to truly "operate computers"Vercel agent-desktop:让 AI Agent 真正「操作电脑」的命令行工具
No screenshots, no pixel matching: It directly reads the accessibility tree of the operating system, transforming desktop applications into JSON structures with ref numbers. AI clicks, inputs, and verifies everything with precision down to the control level. Written in Rust, it works across three platforms – macOS, Windows, and Linux – and topped GitHub's growth chart the week it was released. It marks the beginning of the era of "agents controlling real computers".
不截图、不像素匹配:它直接读取操作系统的无障碍树,把桌面应用变成带 ref 编号的 JSON 结构,AI 点击、输入、验证全部精确到控件级别。Rust 编写,macOS / Windows / Linux 三平台通吃,发布当周登顶 GitHub 增速榜,是「Agent 操控真实计算机」元年的开山之作。
iFixAi: 41,000-star AI Agent Independent Audit Tool, 49 checks 120 seconds to uncover Agent's unauthorized actionsiFixAi:4.1 万星 AI Agent 独立审计工具,49 项检查 120 秒揪出 Agent 的越权操作
Your AI Agent is working for you every day, but do you know if there is anything it shouldn't do? The Python tool that rushed to the top of the GitHub Hot List this week and won the Product Hunt Daily Ranking gives a complete set of answers: Put the Agent into a sandbox that simulates your real business, send out daily requests and requests that intentionally cross the border, and then be scored one by one by an independent referee from different vendors. 2300 + stars this Sunday, from AI Red Team, Governance, Compliance to Philosophy, turning "Agent Can't Be Trusted at All" into a referenced scorecard.
你的 AI Agent 每天都在替你干活,但你知道它有没有偷偷干不该干的吗?这个本周冲上 GitHub 热榜、拿下 Product Hunt 日榜第一的 Python 工具给出了一套完整答案:把 Agent 放进模拟你真实业务的沙盒,发去日常请求和故意越界的请求,再由一个来自不同厂商的独立裁判逐条打分。本周日增 2300+ 星,从 AI 红队、治理、合规到哲学,把「Agent 到底可不可信」变成一张可引用的评分卡。

DeepTutor: 52,000 stars self-hosted AI private education, turning your materials into a learning platform that will ask questions, ask questions, and keep an eye on progressDeepTutor:5.2 万星自托管 AI 私教,把你的资料变成会出题、会追问、会盯进度的学习平台
A collection of essays and textbooks you've never read? This open source project from HKUDS (Hong Kong Big Data Intelligence Laboratory) has made "learning" a complete product: multi-agent collaboration plus rag, turn your documents into a knowledge base, generate structured learning "Book", with quizzes, mastery tracking and learning paths, and remember where you have learned throughout the memory layer. With one-click self-hosting with Docker, your data is all yours.
收藏了一堆论文和教材,从来没读完过?这个来自 HKUDS(港大数据智能实验室)的开源项目把「学习」这件事做成了完整产品:多代理协作加 RAG,把你的文档变成知识库,生成结构化的学习「Book」,配测验、掌握度追踪和学习路径,还有贯穿始终的记忆层记住你学到哪了。可以 Docker 一键自托管,数据全在自己手里。
AI 技能包AI Skill Packs · Give your AI real abilities, out of the box给 AI 装上具体能力,开箱即用

video-shotcraft: A "motion design studio" skill with over 19,200 stars, which lets Claude Code turn screenshots of your real products into promotional videos with cinematic techniques and sound effects.video-shotcraft:19,200+ 星的「动态设计工作室」技能,让 Claude Code 把你的真实产品截图剪成带镜头语言和音效的宣传片
Say "Create a promotional video for my product" to Claude Code or Codex, and this skill developed by a Chinese author will take over the entire process: it will actually capture your page, select camera movements from 124 shot recipe cards, use Remotion to render the animation frame by frame, edit to the beat, add cinematic-quality sound effects, perform frame-by-frame quality checks, and finally deliver an MP4 that can still be further edited. It's not an AI video generator – every frame is drawn by code, so it's clear, on-brand, and fully editable. Since its launch, it has soared to over 19,200 stars, and even the export of CapCut drafts has been taken care of.
给 Claude Code 或 Codex 说一句「给我的产品做个宣传片」,这个中文作者开发的技能会接管全程:真实抓取你的页面、从 124 张镜头配方卡里挑运镜、用 Remotion 逐帧绘制动画、卡节拍剪辑、配电影级音效、逐帧质检,最后交付一条还能继续剪的 MP4。它不是 AI 视频生成器——每一帧都是代码画出来的,所以清晰、贴品牌、完全可编辑。上线以来冲到 19,200+ 星,连剪映草稿导出都做好了。

Cloudflare's Official Open-Source Security Audit Skill: Turn Your AI Agent into a Vulnerability HunterCloudflare 官方开源安全审计技能:把你手上的 AI Agent 变成漏洞挖掘器
This is the "single-repository starting point" of Cloudflare's internal vulnerability discovery system – the seed skills of the harness suite described in an article on the official blog. The six-stage process covers reconnaissance, vulnerability hunting, candidate validation, structured output, independent review, and report generation. It gains 10,000 stars each week, topping the weekly rankings, and a single file of less than 120KB contains 14 attack surface manuals.
这是 Cloudflare 内部漏洞挖掘系统的「单仓库起点」——官方博客亲自撰文介绍的那套 harness 的种子技能。六阶段流程覆盖侦察、猎洞、候选验证、结构化输出、独立复核与报告生成,每周新增一万星登顶周榜,单文件不到 120KB 却装着 14 个攻击面手册。

Unslop design skills: One-click removal of the "obviously fake" look from AI-generated interfacesunslop 设计技能:一键去掉 AI 生成界面的「一眼假」
Purple gradients, glowing effects, emoji tags, shouty copy, uneven spacing – it first lists all the AI elements on the page one by one like an auditor, then extracts the direction from your product itself, and fixes each item like a designer doing a redline review. It supports Next.js, React, Vue, Svelte, Tailwind and pure HTML/CSS, and you can even give it a link to modify a live page.
紫色渐变、发光效果、emoji 标签、喊话式文案、不均匀间距——它先像审计员一样把页面里所有 AI 味逐条列出来,再从你的产品本身提炼方向,像设计师做 redline 评审那样逐项修复。支持 Next.js、React、Vue、Svelte、Tailwind 和纯 HTML/CSS,甚至给它一个链接就能改线上页面。

obsidian-skills: 50,000-star skill pack written by Obsidian CEO to let Claude Code read your entire notebookobsidian-skills:Obsidian CEO 亲写的 5 万星技能包,让 Claude Code 读懂你的整个笔记库
Let AI help you manage the Obsidian library. The biggest problem is that it does not understand the dialect of Obsidian: double chain, embedding, Callout, properties, Bases, JSON Canvas, each is a separate set of syntax. This 50,000 star project is personally maintained by the current CEO of Obsidian, and the six official skills are packaged: from Obsidian flavor Markdown to .base data view to JSON Canvas, all following the agentskills.io standard specification, Claude Code, Codex, OpenCode can be loaded and used.
让 AI 帮你管理 Obsidian 库,最大的坎是它不懂 Obsidian 的方言:双链、嵌入、Callout、属性、Bases、JSON Canvas,每个都是单独一套语法。这个 5 万星项目由 Obsidian 现任 CEO 亲自维护,把六个官方级技能打包好:从 Obsidian 风味 Markdown 到 .base 数据视图再到 JSON Canvas 画布,全部遵循 agentskills.io 标准规范,Claude Code、Codex、OpenCode 装上就能用。
今日正能量保持好奇,持续探索,未来的你会感谢今天的开始。挑一个好工具上手 →广告免费下载专区真实文件,点击即下,无需注册
claude-obsidian:让 Claude Code 和 Obsidian 变成一对「本地优先」知识工作搭档,任何资料丢进去都变成带引用链接的笔记
claude-obsidian v3.0.1 完整插件包
browser-use:11.9 万星的浏览器自动化王者发布 1.0 稳定版,AI 操控浏览器这件事终于有了事实标准
browser-use v1.0.4 官方 wheel 安装包
video-shotcraft:19,200+ 星的「动态设计工作室」技能,让 Claude Code 把你的真实产品截图剪成带镜头语言和音效的宣传片
video-shotcraft 精华技能包(SKILL.md + 全套镜头配方库)
微软官方开源 TokenEconomizer:不改 Agent 一行代码,实测编码成本直降 30.9%
TokenEconomizer 完整插件源码包(Claude Code / Codex / Copilot CLI / Pi)
Vercel agent-desktop:让 AI Agent 真正「操作电脑」的命令行工具
agent-desktop v1.3.0 Windows x64 原生可执行文件
Cloudflare 官方开源安全审计技能:把你手上的 AI Agent 变成漏洞挖掘器
security-audit 完整技能包(SKILL.md + 14 攻击面手册 + 零依赖校验器)
编辑精选值得先看的几篇

Claude-Obsidian: Turn Claude Code and Obsidian into a "local-first" knowledge work partnership. Any information you put in becomes a note with a reference link.claude-obsidian:让 Claude Code 和 Obsidian 变成一对「本地优先」知识工作搭档,任何资料丢进去都变成带引用链接的笔记
Do you have two hundred unread articles piled up in your favourites and dozens of PDFs sitting in your downloads folder, and you want AI to help you process them but are always worried that it will make up sources? This open-source project with over 16,400 stars directly connects Claude Code (or any host compatible with Agent Skills) to the Obsidian vault: you feed it any source, and it generates interlinked Markdown notes with each conclusion citing its source. When you ask questions, the answers are strictly based on the evidence already in the vault, and if it can't find an answer, it simply says so. It runs entirely locally, and your data never leaves your computer.
收藏夹里堆了两百篇没读的文章、下载目录里躺着几十份 PDF,想让 AI 帮你消化却总担心它瞎编出处?这个 16,400+ 星的开源项目把 Claude Code(或任何兼容 Agent Skills 的宿主)和 Obsidian 库直接接通:你把任何来源丢进去,它产出互相链接、每个结论都标明出处的 Markdown 笔记;提问时回答也严格基于库里已有证据,不给答案就直说找不到。全程本地运行,数据不出你的电脑。

browser-use: The browser automation champion with 119,000 stars releases its stable version 1.0. AI-controlled browsers finally have a de facto standard.browser-use:11.9 万星的浏览器自动化王者发布 1.0 稳定版,AI 操控浏览器这件事终于有了事实标准
Enabling an AI Agent to open a browser, click buttons, fill out forms, and scrape data is something that browser-use has achieved as the most widely cited open-source implementation in the entire field – with 119,000 stars on GitHub. Over the past year, almost all "AI-controlled browser" projects have been built on top of it. The project has now released the 1.0.4 stable version: rewritten based on Pydantic AI, with a persistent Python REPL running directly on the native CDP protocol. It can take over your local Chrome and also supports cloud browsers, with both speed and stability improved.
让 AI Agent 自己开浏览器、点按钮、填表单、抓数据,这件事 browser-use 已经做成了整个领域引用最多的开源实现——GitHub 11.9 万星,一年来几乎所有「AI 操控浏览器」项目都在它之上搭建。现在项目发布 1.0.4 稳定版:基于 Pydantic AI 重写,一条持久的 Python REPL 直接跑在原生 CDP 协议上,可接管你本地的 Chrome,也支持云端浏览器,速度和稳定性都上了一个台阶。

video-shotcraft: A "motion design studio" skill with over 19,200 stars, which lets Claude Code turn screenshots of your real products into promotional videos with cinematic techniques and sound effects.video-shotcraft:19,200+ 星的「动态设计工作室」技能,让 Claude Code 把你的真实产品截图剪成带镜头语言和音效的宣传片
Say "Create a promotional video for my product" to Claude Code or Codex, and this skill developed by a Chinese author will take over the entire process: it will actually capture your page, select camera movements from 124 shot recipe cards, use Remotion to render the animation frame by frame, edit to the beat, add cinematic-quality sound effects, perform frame-by-frame quality checks, and finally deliver an MP4 that can still be further edited. It's not an AI video generator – every frame is drawn by code, so it's clear, on-brand, and fully editable. Since its launch, it has soared to over 19,200 stars, and even the export of CapCut drafts has been taken care of.
给 Claude Code 或 Codex 说一句「给我的产品做个宣传片」,这个中文作者开发的技能会接管全程:真实抓取你的页面、从 124 张镜头配方卡里挑运镜、用 Remotion 逐帧绘制动画、卡节拍剪辑、配电影级音效、逐帧质检,最后交付一条还能继续剪的 MP4。它不是 AI 视频生成器——每一帧都是代码画出来的,所以清晰、贴品牌、完全可编辑。上线以来冲到 19,200+ 星,连剪映草稿导出都做好了。

Microsoft's Official Open-Source TokenEconomizer: Without Changing a Single Line of Code in the Agent, Actual Coding Costs Are Reduced by 30.9%微软官方开源 TokenEconomizer:不改 Agent 一行代码,实测编码成本直降 30.9%
Presented by UC Berkeley in collaboration with Microsoft Research: Learn three techniques from your past sessions – micro-skills to remember repository rules, eliminate tools that have never been used, and stop failure loops that go round and round. SWE-Bench tested 277 real-world tasks, and all 30 "repository × model" combinations were cheaper, with a slight increase in task success rate. Supports Claude Code, Codex, GitHub Copilot CLI, and Pi.
UC Berkeley 联手微软研究院出品:从你跑过的历史会话里学三招——微技能记住仓库规矩、裁掉从没用的工具、拦住原地打转的失败循环。SWE-Bench 277 个真实任务实测,30 个「仓库×模型」组合全部更便宜,任务成功率还略有提升。支持 Claude Code、Codex、GitHub Copilot CLI、Pi。

Vercel agent-desktop: A command-line tool that enables AI agents to truly "operate computers"Vercel agent-desktop:让 AI Agent 真正「操作电脑」的命令行工具
No screenshots, no pixel matching: It directly reads the accessibility tree of the operating system, transforming desktop applications into JSON structures with ref numbers. AI clicks, inputs, and verifies everything with precision down to the control level. Written in Rust, it works across three platforms – macOS, Windows, and Linux – and topped GitHub's growth chart the week it was released. It marks the beginning of the era of "agents controlling real computers".
不截图、不像素匹配:它直接读取操作系统的无障碍树,把桌面应用变成带 ref 编号的 JSON 结构,AI 点击、输入、验证全部精确到控件级别。Rust 编写,macOS / Windows / Linux 三平台通吃,发布当周登顶 GitHub 增速榜,是「Agent 操控真实计算机」元年的开山之作。

Cloudflare's Official Open-Source Security Audit Skill: Turn Your AI Agent into a Vulnerability HunterCloudflare 官方开源安全审计技能:把你手上的 AI Agent 变成漏洞挖掘器
This is the "single-repository starting point" of Cloudflare's internal vulnerability discovery system – the seed skills of the harness suite described in an article on the official blog. The six-stage process covers reconnaissance, vulnerability hunting, candidate validation, structured output, independent review, and report generation. It gains 10,000 stars each week, topping the weekly rankings, and a single file of less than 120KB contains 14 attack surface manuals.
这是 Cloudflare 内部漏洞挖掘系统的「单仓库起点」——官方博客亲自撰文介绍的那套 harness 的种子技能。六阶段流程覆盖侦察、猎洞、候选验证、结构化输出、独立复核与报告生成,每周新增一万星登顶周榜,单文件不到 120KB 却装着 14 个攻击面手册。

Unslop design skills: One-click removal of the "obviously fake" look from AI-generated interfacesunslop 设计技能:一键去掉 AI 生成界面的「一眼假」
Purple gradients, glowing effects, emoji tags, shouty copy, uneven spacing – it first lists all the AI elements on the page one by one like an auditor, then extracts the direction from your product itself, and fixes each item like a designer doing a redline review. It supports Next.js, React, Vue, Svelte, Tailwind and pure HTML/CSS, and you can even give it a link to modify a live page.
紫色渐变、发光效果、emoji 标签、喊话式文案、不均匀间距——它先像审计员一样把页面里所有 AI 味逐条列出来,再从你的产品本身提炼方向,像设计师做 redline 评审那样逐项修复。支持 Next.js、React、Vue、Svelte、Tailwind 和纯 HTML/CSS,甚至给它一个链接就能改线上页面。
stop-that-shit: Chinese open source "Mission Boundary Guard", a red stop seal to shut up when the AI is donestop-that-shit:中文开源「任务边界守卫」,一个红色 STOP 印章让 AI 干完正事就闭嘴
You only ask the Agent to export one file, and it calculates the SHA-256 that no one reads, saying "just in case"; you let it do only Review, and it finds the problem and starts to change the code. This Chinese open source plugin, whose name is even straightforward enough, uses a set of Skill plus Guard hooks to stop the four types of redundant actions of AI in real time in mainstream hosts such as Codex, Claude Code, OpenCode, and Hermes Agent CLI - range expansion, useless defense, intention to cross the border, and task rotation. The official benchmark can save about 65% of the nonsense output tokens, and the technical content is not discounted at all.
你只让 Agent 导出一个文件,它顺手算了份没人读取的 SHA-256,说是「以防万一」;你让它只做 Review,它找到问题就开始改代码。这个连名字都够直白的中文开源插件,用一套 Skill 加 Guard 钩子,在 Codex、Claude Code、OpenCode、Hermes Agent CLI 等主流宿主里实时拦下 AI 的四类多余动作——范围膨胀、无用防御、意图越界、任务打转,官方基准能省约 65% 的废话输出 token,而且技术内容一点不打折。
iFixAi: 41,000-star AI Agent Independent Audit Tool, 49 checks 120 seconds to uncover Agent's unauthorized actionsiFixAi:4.1 万星 AI Agent 独立审计工具,49 项检查 120 秒揪出 Agent 的越权操作
Your AI Agent is working for you every day, but do you know if there is anything it shouldn't do? The Python tool that rushed to the top of the GitHub Hot List this week and won the Product Hunt Daily Ranking gives a complete set of answers: Put the Agent into a sandbox that simulates your real business, send out daily requests and requests that intentionally cross the border, and then be scored one by one by an independent referee from different vendors. 2300 + stars this Sunday, from AI Red Team, Governance, Compliance to Philosophy, turning "Agent Can't Be Trusted at All" into a referenced scorecard.
你的 AI Agent 每天都在替你干活,但你知道它有没有偷偷干不该干的吗?这个本周冲上 GitHub 热榜、拿下 Product Hunt 日榜第一的 Python 工具给出了一套完整答案:把 Agent 放进模拟你真实业务的沙盒,发去日常请求和故意越界的请求,再由一个来自不同厂商的独立裁判逐条打分。本周日增 2300+ 星,从 AI 红队、治理、合规到哲学,把「Agent 到底可不可信」变成一张可引用的评分卡。

obsidian-skills: 50,000-star skill pack written by Obsidian CEO to let Claude Code read your entire notebookobsidian-skills:Obsidian CEO 亲写的 5 万星技能包,让 Claude Code 读懂你的整个笔记库
Let AI help you manage the Obsidian library. The biggest problem is that it does not understand the dialect of Obsidian: double chain, embedding, Callout, properties, Bases, JSON Canvas, each is a separate set of syntax. This 50,000 star project is personally maintained by the current CEO of Obsidian, and the six official skills are packaged: from Obsidian flavor Markdown to .base data view to JSON Canvas, all following the agentskills.io standard specification, Claude Code, Codex, OpenCode can be loaded and used.
让 AI 帮你管理 Obsidian 库,最大的坎是它不懂 Obsidian 的方言:双链、嵌入、Callout、属性、Bases、JSON Canvas,每个都是单独一套语法。这个 5 万星项目由 Obsidian 现任 CEO 亲自维护,把六个官方级技能打包好:从 Obsidian 风味 Markdown 到 .base 数据视图再到 JSON Canvas 画布,全部遵循 agentskills.io 标准规范,Claude Code、Codex、OpenCode 装上就能用。

watermarks-remover: 25,000-star open source plugin to clean up the AI watermarks Claude, Gemini, OpenAI buried in your fileswatermarks-remover:2.5 万星开源插件,把 Claude、Gemini、OpenAI 埋在你文件里的 AI 水印一次性剥干净
The documents you ask AI to help write and the images you generate may contain invisible "ID cards": invisible Unicode characters, statistical text watermarks, C2PA/EXIF metadata, all quietly marked "This content is generated by AI". This open source project in October studied the traceability marks of major manufacturers thoroughly, added a proxy skill to the service implemented by a pure Python standard library, and cleaned them on your own computer - no online uploads, no touching other people's content, and focused on "only handling your own things".
你让 AI 帮忙写的文档、生成的图片,里面可能藏着看不见的「身份证」:不可见的 Unicode 字符、统计型文本水印、C2PA/EXIF 元数据,都在悄悄标注「这段内容由 AI 生成」。这个 10 月蹿红的开源项目把各大厂商的溯源标记研究了个透,用一个纯 Python 标准库实现的服务加一个代理技能,在你自己的电脑上把它们清干净——不联网上传、不碰别人的内容,主打一个「只处理你自己的东西」。

taste-skill: 106,000-star design taste skill pack, so that the page handed over by AI is no longer "fake at a glance"taste-skill:10.6 万星设计品味技能包,让 AI 交出来的页面不再「一眼假」
AI-written webpages always have a familiar pipelined flavor: Inter font, purple-blue gradient, card holder cards. What this skill pack does is big, small, and not small - add "design taste" to the coding agent: first set the art direction, then find the real photo with pictures from the open gallery, and finally put the page into the real browser for self-check according to the three widths. Check out an "AI taste" and hit back to redo. Built on research from 247 curated websites, it is currently one of the most installed design skills and is supported by Vercel's official open source initiative.
AI 写的网页总有一股熟悉的流水线味:Inter 字体、紫蓝渐变、卡片套卡片。这个技能包做的事情说大不大、说小不小——给编码代理装上「设计品味」:先定艺术方向,再从开放图库找真实照片配图,最后把页面放进真浏览器里按三种宽度自检,检出一项「AI 味」就打回去重做。它基于对 247 个精选网站的研究构建,是当前安装量最高的设计类技能之一,Vercel 官方开源计划也在支持它。

DeepTutor: 52,000 stars self-hosted AI private education, turning your materials into a learning platform that will ask questions, ask questions, and keep an eye on progressDeepTutor:5.2 万星自托管 AI 私教,把你的资料变成会出题、会追问、会盯进度的学习平台
A collection of essays and textbooks you've never read? This open source project from HKUDS (Hong Kong Big Data Intelligence Laboratory) has made "learning" a complete product: multi-agent collaboration plus rag, turn your documents into a knowledge base, generate structured learning "Book", with quizzes, mastery tracking and learning paths, and remember where you have learned throughout the memory layer. With one-click self-hosting with Docker, your data is all yours.
收藏了一堆论文和教材,从来没读完过?这个来自 HKUDS(港大数据智能实验室)的开源项目把「学习」这件事做成了完整产品:多代理协作加 RAG,把你的文档变成知识库,生成结构化的学习「Book」,配测验、掌握度追踪和学习路径,还有贯穿始终的记忆层记住你学到哪了。可以 Docker 一键自托管,数据全在自己手里。

Knowledge Work Plugins: Anthropic official open source 11 career plugins, sales, finance, legal, data, PM each have a set of ready-made workflowsKnowledge Work Plugins:Anthropic 官方开源 11 个职业插件,销售、财务、法务、数据、PM 各有一套现成工作流
Most of the Claude plug-ins on the market are spelled by individual authors, and the quality is uneven. This set is different - it comes from Anthropic itself, 32,000 stars, and breaks down the professional process of 11 functional positions into plugins: each plugin packages skills, external tool connectors, slash commands, and subagents. After installing it, Claude is no longer a "know-it-all" general assistant, but a colleague who knows your rules: the sales plug-in will prepare the call at your rhythm, the financial plug-in knows what step to do first, and the data plug-in must first self-test after writing SQL. All Markdown and JSON, no code, no build steps, change is change the text.
市面上的 Claude 插件大多是个人作者拼的,质量参差。这一套不一样——它出自 Anthropic 自己,3.2 万星,把 11 个职能岗位的专业流程拆成插件:每个插件里打包了技能(skills)、外部工具连接器(connectors)、斜杠命令和子代理。装上之后 Claude 不再是一个「什么都懂一点」的通用助手,而是懂你这行规矩的同事:销售插件会按你的节奏准备通话,财务插件知道月结该先做哪一步,数据插件写完 SQL 还要先自检。全部是 Markdown 和 JSON,没有代码、没有构建步骤,改起来就是改文本。

Open Duet: Raise a team of AI colleagues on your own computer or server, with a name, a position, and a memory, and you need to nod before you startOpen Duet:在自己电脑或服务器上养一支 AI 同事团队,有名字、有岗位、有记忆,动手前还要你点头
The logic of most proxy tools is “one person facing a chat box”. Open Duet has a different model: it is an open source workspace, you can create a bunch of AI colleagues with names, assign each position, specify the workspace, assign the model, set the memory, and let them assign each other - leading the agent to meet the needs, disassemble the tasks, hand the work to the expert agent, and then return to the chat. The data stays on your own machine or server, and when it needs approval is up to you. 6,100 + stars, rushing to the GitHub Trends this week.
大多数代理工具的逻辑是「一个人对着一个聊天框」。Open Duet 换了个模型:它是个开源工作区,你可以创建一堆有名字的 AI 同事,给每个安排岗位、指定工作区、分配模型、设置记忆,让他们互相派活——主导代理接需求、拆任务,把活交给专家代理,结果回到聊天里。数据留在你自己的机器或服务器上,什么时候需要人批准由你决定。6,100+ 星,本周冲上 GitHub 趋势榜。

book-to-skill: Compressing a technical book into a skill pack, and only reading the correct chapter when asking questions, saves 24-51 times the token than cramming the entire book into the contextbook-to-skill:把一本技术书压成技能包,提问时只读对的那一章,比把整本书塞进上下文省 24–51 倍 token
I bought a good book and read it again. After three months, I even forgot that Chapter 7 existed. It is also not cost-effective for an agent to read a PDF directly - it has to revisit the catalog, turn it over, and process the entire book again every time. The author of this cost project is called "Discovery Circular Tax". book-to-skill (36,000 stars, mit) does this all at once: extract the book or a batch of documents into structured skills - mental models, decision-making rules, anti-patterns, chapter-by-chapter documents, and then you ask what, the agent only loads the relevant chapter, and the answer comes from the original text rather than illusions.
买过的好书,读一遍,三个月后连第七章存在过都忘了。让代理直接读 PDF 也不划算——它每次都要重新翻目录、回翻、把整本书再处理一遍,这套开销项目作者叫「发现循环税」。book-to-skill(3.6 万星,MIT)把这件事一次性做完:把书或一整批文档抽成结构化的技能——心智模型、决策规则、反模式、按章拆分的文件,之后你问什么,代理只加载相关那一章,答案来自原文而不是幻觉。
今日正能量学一点,用一点;今天种下的种子,会长成明天的能力。去免费下载专区 →广告




