Great AI tools,
right in your hands.把好用的 AI，
直接放到你手里。

Byte Seed team found DeepSeek "pumping" reason, long context probability can drift字节 Seed 团队发现 DeepSeek“抽风”原因，长上下文可能性能漂移

Synopsys intends to seek cooperation with AI laboratories in China to accelerate chip design in China新思科技拟寻求与中国 AI 实验室合作，在华加速芯片设计

Google Cloud Releases Gemini Agent; Manus Official Announces $500 Million in Financing; Xiaopeng Launches Robotaxi Taxi Applet谷歌云发布 Gemini Agent；Manus 官宣五亿美元融资；小鹏上线 Robotaxi 打车小程序

California seeks to ban “man-machine fights”, accusing the state of illegally staging matches加州试图禁止“人机格斗赛”，指控已举办的对决赛事非法

The chatbot encouraged users to self-harm and starve, and Character.AI was accused by the state of Kentucky of "defective products"聊天机器人怂恿用户自残、挨饿，Character.AI 遭肯塔基州指控“产品存在缺陷”

OpenAI launches GPT-6.1 Sol Ultrafast Edition: 6x price, up to 8x speedOpenAI 推出 GPT-6.1 Sol Ultrafast 版：6 倍价格、最高 8 倍速度

Alibaba Tsai Chongxin: In five years, AI will be integrated into every aspect of society like today's Internet阿里巴巴蔡崇信：五年后 AI 将如今天的互联网，融入社会的每个环节

3 fired OpenAI researchers send open letter saying company practices could trigger 'chilling effect'3 名遭解雇 OpenAI 研究员发公开信，称公司做法或引发“寒蝉效应”

Graphics Tutorial: Installing MCP for ChatGPT/DeepSeek Web, from plugin to first local tool图文教程：给 ChatGPT / DeepSeek 网页版装上 MCP，从装插件到跑通第一个本地工具

This tutorial will help you turn the web-based AI you use every day into a working assistant: first build the MCP SuperAssistant extension with the source code, then start an MCP service locally, and finally go back to the ChatGPT web version and click the button to let the AI actually read the file on your machine. No API Key is required throughout, and no AI platform settings need to be changed.

这篇教程带你把自己天天在用的网页版 AI 变成能真正干活的助手：先用源码构建 MCP SuperAssistant 扩展，再在本地起一个 MCP 服务，最后回到 ChatGPT 网页版里点一下按钮，让 AI 真的读到你本机上的文件。全程不需要 API Key，也不需要改任何 AI 平台设置。

Text tutorial: Save Kanban from installation to reading by giving Claude Code and Codex save tokens with RTK图文教程：用 RTK 给 Claude Code 与 Codex 省 token，从安装到读懂节省看板

Many people think that saving money depends on replacing a smaller model. In fact, a more effective step is not to feed the noise to the model first. This tutorial will guide you to load RTK into the AI programming assistant: do a global initialization first, then verify whether the hook is effective, then open the save Kanban board to compare before and after, and finally explain which cases it can not help, to avoid the effect of misjudgment.

很多人以为省钱要靠换更小的模型，其实更有效的一步是先别把噪声喂给模型。这篇教程带你把 RTK 装进 AI 编程助手：先做一次全局初始化，再验证钩子是否生效，然后打开节省看板对照前后来看，最后讲清楚哪些情况下它帮不上忙，避免误判效果。

Image and text tutorial: 5 minutes to pack the tree and cut the webpage and video into Obsidian (including AI question configuration)图文教程：5 分钟装好乔木剪藏，把网页和视频剪进 Obsidian（含 AI 提问配置）

From downloading the installation package to cutting the first article, you only need to press the keyboard three times in a row. This tutorial walks through it according to the real interface screenshots: installing plug-ins, matching AI, watching videos for bilingual notes, cutting to Obsidian with one click, and finally giving a shortcut key cheat sheet. By doing this, you can turn "favorite gray" into a searchable note library.

从下载安装包到剪下第一篇文章，全程只需要连着按三下键盘。这篇教程按真实界面截图走一遍：装插件、配 AI、看视频做双语笔记、一键剪藏到 Obsidian，最后给出快捷键速查表，照着做就能把「收藏夹吃灰」变成可检索的笔记库。

Teletext Tutorial: Deploy OpenCut locally and set up your own free editing workbench in a dozen minutes图文教程：本地部署 OpenCut，十几分钟搭一套属于自己的免费剪辑工作台

Don't want to pay for the annual cost of editing software anymore, and don't want to pass customer material to someone else's cloud? This tutorial shows you how to use the source code to run the classic version of OpenCut locally: install the environment, pull the code, load the dependency, start the service, then build the project, upload the material, cut the first video, and finally explain how to deploy it to your own domain name for the team to share.

不想再为剪辑软件年费买单、也不想把客户素材传到别人的云上？这篇教程带你用源码在本地把 OpenCut 经典版跑起来：装环境、拉代码、装依赖、启动服务，然后建工程、上素材、剪出第一段成片，最后讲清楚怎么把它部署到自己的域名上给团队共用。

MCP SuperAssistant browser plugin: let ChatGPT, Gemini, DeepSeek web directly call MCP tools (mit open source)MCP SuperAssistant 浏览器插件：让 ChatGPT、Gemini、DeepSeek 网页版直接调用 MCP 工具（MIT 开源）

AI on the web has always had a hard time - only chatting, not touching your local files, databases, and business systems. MCP SuperAssistant uses a lightweight Chrome extension to cram the Model Context Protocol into more than a dozen mainstream webpages such as ChatGPT, Gemini, DeepSeek, Kimi, Grok, Perplexity, etc.: The tool call that appears in the AI answer will be automatically recognized as a one-click execution button, and the native MCP service will actually run through it with one click, and the result will be automatically backfilled into the conversation. Mit is open source, front-end-only, and does not transit through any cloud.

网页版 AI 一直有个硬伤——只能聊天，碰不到你的本地文件、数据库和业务系统。MCP SuperAssistant 用一个轻量的 Chrome 扩展把 Model Context Protocol 塞进 ChatGPT、Gemini、DeepSeek、Kimi、Grok、Perplexity 等十几个主流网页端：AI 回答里出现的工具调用会被自动识别成一键执行的按钮，点一下就由本机 MCP 服务真正跑完，结果自动回填进对话。MIT 开源、纯前端、不经任何云端中转。

Tree Clipping Qiaomu Clipper: Clip webpages and videos into Obsidian's open source plugin and ask AI while reading (rushing to GitHub Trends for three days)乔木剪藏 Qiaomu Clipper：把网页和视频剪进 Obsidian 的开源插件，边读边问 AI（发布三天冲上 GitHub 趋势榜）

A Chrome extension made by a Chinese team, specializing in "collecting gray": when you read a good article, press aaa to enter a clean reading page, cross the line to ask questions on the spot AI; watch YouTube, B site video, press aaa to open the bilingual subtitle toolbar, click the timestamp to jump back to that sentence; when you see half of the sudden ideas, press N to write a small card, along with the source and time point, automatically added to today's Obsidian diary. Mit is open source, the data is left on the machine, and no model can be used as a pure high-quality clipping tool.

一个中文团队做的 Chrome 扩展，专治「收藏夹吃灰」：读到好文章按 aaa 进入干净阅读页，划线就能就地提问 AI；看 YouTube、B 站视频按 aaa 打开双语字幕工具栏，点时间戳直接跳回那句话；看到一半突然有想法，按 N 写一张小卡片，连同来源与时间点自动追加进今天的 Obsidian 日记。MIT 开源、数据留在本机，不配任何模型也能当纯粹的高质量剪藏工具用。

Read Frog: Open source bilingual control reading plugin, webpage, YouTube subtitles, e-books together (nearly 10,000 stars)Read Frog 陪读蛙：开源双语对照阅读插件，网页、YouTube 字幕、电子书一起翻（近万 Star）

An open-source browser plug-in that turns "reading foreign websites" into "learning foreign languages": full-page bilingual contrast, transcription translation with explanation, YouTube bilingual subtitles, e-book translation, support for more than 20 models such as OpenAI, DeepSeek, Gemini, Ollama, and a free Google/Microsoft translation engine. Bulk requests can save up to 70% on API fees, while local connections to Ollama are completely free and data is local.

一款把「读外网」变成「学外语」的开源浏览器插件：整页双语对照、划词翻译带讲解、YouTube 双语字幕、电子书翻译一站搞定，支持 OpenAI、DeepSeek、Gemini、Ollama 等 20 多家模型，也内置免费的谷歌/微软翻译引擎。批量请求合并最高能省 70% 的 API 费用，本地接 Ollama 则完全免费、数据不出本机。

MCP for Blender: Let AI directly take over Blender and create 3D scenes in one sentence (30,000 Star open source plug-in)MCP for Blender：让 AI 直接接管 Blender，一句话建出 3D 场景（3 万 Star 开源插件）

It has 30,000 stars on GitHub and has long dominated the 3D and AI tool list. It is a Blender plug-in plus an MCP server: after installing it, you say in Chinese in Claude, Cursor, and Codex "Make a dungeon scene, with a dragon guarding the treasure chest." The AI ​​will model, add materials, lighting, and adjust the camera in Blender itself, and it can also send back screenshots at any time to check whether it is done correctly. In addition to do-it-yourself modeling, it can also directly adjust the free model libraries of Poly Haven, Sketchfab, and Poly Pizza, as well as Vincent 3D interfaces such as Hyper3D and Tencent Hunyuan 3D.

GitHub 3 万 Star、长期霸榜 3D 与 AI 工具榜。它是一个 Blender 插件加一个 MCP 服务端：装好之后，你在 Claude、Cursor、Codex 里用中文说「做一个地下城场景，有条龙守着宝箱」，AI 就自己在 Blender 里建模、加材质、布光、调相机，还能随时回传截图核对做得对不对。除了自己动手建模，它还能直接调 Poly Haven、Sketchfab、Poly Pizza 的免费模型库，以及 Hyper3D、腾讯混元 3D 这类文生 3D 接口。

RTK (Rust Token Killer): Cut the command output read by the AI assistant by up to 90%, single file Rust binary, 80,000 starsRTK（Rust Token Killer）：把 AI 助手读到的命令输出砍掉最多九成，单文件 Rust 二进制、8 万星

People who write code with Claude Code, Codex, and Cursor are distressed by tokens: a git status, a test failure log, thousands of tokens are poured into the context, the money is spent, and the window is full. RTK is an "output scissor" written in Rust - it inserts a layer of proxy between the command and the model, and filters, groups, truncates, deduplicates and hands over the output of more than 100 commands such as ls, grep, git, test, build, Docker, kubectl, etc. to the model. Officially, it can reduce the bash output by up to 90%, with a single overhead of less than 10 milliseconds.

用 Claude Code、Codex、Cursor 写代码的人都心疼过 token：一条 git status、一次测试失败日志，动辄几千上万 token 灌进上下文，钱花了、窗口也被占满。RTK 是一把用 Rust 写的「输出剪刀」——它在命令与模型之间插一层代理，把 ls、grep、git、测试、构建、Docker、kubectl 等 100 多个命令的输出做过滤、分组、截断、去重后再交给模型，官方称最高可减少约九成的 bash 输出，单次开销低于 10 毫秒。

OpenCut: 93,000 Star Open Source Clipper, Free Clipping/CapCut Alternative Running LocallyOpenCut：9.3 万 Star 的开源剪辑器，本地运行的免费剪映 / CapCut 替代方案

Editing software is getting more and more expensive, and the export is also watermarked? OpenCut is currently the most popular video clip project in the open source circle. With 93,000 Star and mit licenses, you can run a complete timeline clip in the browser, and the material is not uploaded to a third-party cloud. It turned the "timeline, transition, special effects, text overlay, audio mixing" set of professional tools into a web application, self-hosted one is equivalent to having your own clipping, and easily solved the problem of privacy and subscription fees together.

剪辑软件越来越贵、导出还带水印？OpenCut 是目前开源圈热度最高的视频剪辑项目，9.3 万 Star、MIT 许可，浏览器里就能跑完整时间线剪辑，素材不上传第三方云端。它把「时间线、转场、特效、文字叠加、音频混音」这套专业工具做成了网页应用，自托管一份就等于拥有了自己的剪映，还顺手把隐私和订阅费问题一起解决了。

e2e: Open Source AI Testing Framework for Apps Powered by Natural Language, # 1 on GitHub Trends (6300 + Star)e2e：用自然语言驱动 App 的开源 AI 测试框架，GitHub 趋势日榜第一（6300+ Star）

TesterArmy's next-generation end-to-end test framework: use one sentence to describe the goal (such as "upgrade the workspace to the Pro package"), the agent will go to the page, fill out the form, go through the process, and then verify the result with the locator and assertion in the same test. The web and mobile end are omnipresent, and the built-in agent steps to record and playback - the application does not change and does not burn the model repeatedly, and ordinary test cases do not need to call AI at all.

TesterArmy 出品的下一代端到端测试框架：用一句人话描述目标（比如「把工作区升级到 Pro 套餐」），智能体自己去点页面、填表单、走完流程，再用同一条测试里的定位器和断言验证结果。Web 和移动端通吃，自带智能体步骤录制回放——应用没变就不重复烧模型，普通测试用例更是完全不需要调用 AI。

Strix: an open source AI penetration testing tool that allows AI to run applications like a real hacker, find vulnerabilities and provide repair patchesStrix：开源 AI 渗透测试工具，让 AI 像真黑客一样把应用跑起来、找出漏洞并给出修复补丁

An autonomous penetration testing agent with 66,000 stars. The biggest difference between it and traditional scanners is "verification": AI does not just report a suspicious point, but actually runs the application, constructs a reproducible PoC to prove the existence of the vulnerability, and then provides patches and compliance-level reports. The entire attack toolchain—HTTP interception proxy, browser exploits, shell, Python vulnerability runtime, reconnaissance and OSINT, SAST+DAST—is available out of the box.

6.6 万 Star 的自主渗透测试智能体。它和传统扫描器最大的区别是「验证」：AI 不是只报一个可疑点，而是真的把应用跑起来、构造可复现的 PoC 证明漏洞存在，再给出补丁与合规级报告。整套攻击工具链——HTTP 拦截代理、浏览器利用、Shell、Python 漏洞运行时、侦察与 OSINT、SAST+DAST——都是开箱即用。

Understand Anything: 85,000-star code understanding skill pack that turns any code base into a conversational knowledge graphUnderstand Anything：8.5 万星的代码理解技能包，把任意代码库变成可以对话的知识图谱

The most time-consuming thing to take on an unfamiliar project is not to write code, but to wonder “how did this thing get up?” Understand Anything is a skill pack for the AI programming assistant: an instruction scans the entire repository, extracts each file, function, class, and dependency, and generates an interactive knowledge graph; then you can ask it in natural language "how the payment process goes", let it analyze which modules your current changes will affect, and even generate a newcomer onboarding guide. It supports more than a dozen platforms such as Claude Code, Codex, Cursor, Copilot, Gemini CLI, etc., and mit is open source.

接手一个陌生项目最费时间的不是写代码，而是搞不清「这东西到底怎么跑起来的」。Understand Anything 是给 AI 编程助手装的一个技能包：一条指令扫完整个仓库，抽取每个文件、函数、类与依赖关系，生成一张可交互的知识图谱；随后你可以用自然语言问它「支付流程是怎么走的」、让它分析你当前改动会波及哪些模块、甚至生成新人上手指南。支持 Claude Code、Codex、Cursor、Copilot、Gemini CLI 等十余个平台，MIT 开源。

Replica Skill Pack: 11 Chain Claude Skills, from reverse to “Cleanroom” replica of an app (800 + Stars in four days)Replica 技能包：11 个链式 Claude 技能，从逆向到上线「干净室」复刻一款 App（四天冲上 800+ Star）

Seeing a subscription app that says, “I can do this too,” but it only takes a few weeks to sort through its pages, processes, and data models? Replica divided this into 11 Claude skills that can be executed sequentially: reverse the screen, process, component and data model with public pages and screenshots, plan the technology stack and database, rebuild the design system, develop screen by screen, connect to login and pay, run the test again, and finally read the real bad reviews to change the differences and selling points until the deployment is launched. Mit license, free, no API Key, the tool only uses the Python standard library.

看到一款订阅制 App 觉得「这个我也能做」，但光是梳理它的页面、流程和数据模型就要几周？Replica 把这件事拆成 11 个可以顺序执行的 Claude 技能：用公开页面和截图逆向出屏幕、流程、组件与数据模型，规划技术栈与数据库，重建设计系统，逐屏开发，接上登录支付，跑一遍测试，最后再读真实差评改出差异化和卖点，直到部署上线。MIT 许可、免费、无需 API Key，工具只用 Python 标准库。

gstack: Claude Code virtual engineering team open sourced by YC President Garry Tan, 23 characters at a time (135,000 Stars)gstack：YC 总裁 Garry Tan 开源的 Claude Code 虚拟工程团队，23 个角色一次装齐（13.5 万 Star）

Garry Tan loaded his 20-year engineering process into Claude Code: a CEO who re-examined the product, an engineering manager who locked down the architecture, a designer who specialized in AI spam code, a QA supervisor who really opened the browser for acceptance, a security officer who ran the OWASP + stride audit, and a release engineer who was responsible for the release - 23 expert roles, 8 tools, all with slash commands and Markdown, and the mit license was taken away for free. He claims that with this set of things, the output of logic code is 810 times faster than in 2013.

Garry Tan 把自己二十年的工程流程装进了 Claude Code：一个重新审视产品的 CEO、一个锁死架构的工程经理、一个专抓 AI 垃圾代码的设计师、一个真开浏览器做验收的 QA 主管、一个跑 OWASP + STRIDE 审计的安全官，外加负责发布的发布工程师——23 个专家角色、8 件工具，全是斜杠命令和 Markdown，MIT 许可免费拿走。他自称靠这套东西，逻辑代码产出速度达到了 2013 年的 810 倍。

Marketing Skills: 50+ marketing skills package to turn an AI programming assistant into a marketing teamMarketing Skills：50+ 营销技能包，把 AI 编程助手变成一支营销团队

53,000 Stars, a newcomer on GitHub trend list this week. It writes a complete set of marketing strategies such as conversion optimization, copywriting, SEO, delivery, data analysis, and growth into 50+ Markdown skill files loaded on demand, and feeds them to Claude Code, Codex, Cursor and other agents that support the Agent Skills specification. The core is a basic product-marketing skill: first explain your product, audience, and positioning clearly once, and other skills will automatically reference it, and the output will no longer be a general cliché.

5.3 万 Star、本周 GitHub 趋势榜新贵。它把转化优化、文案、SEO、投放、数据分析、增长等一整套营销打法，写成 50+ 个按需加载的 Markdown 技能文件，喂给 Claude Code、Codex、Cursor 等支持 Agent Skills 规范的智能体。核心是一个 product-marketing 基础技能：先把你产品、受众、定位讲清楚一次，其他技能都会自动引用它，输出就不再是泛泛而谈的套话。