AI×设计：openpencil（MIT，本机读写 .fig/.pen，CLI @open-pencil/cli、MCP @open-pencil/mcp、Vue SDK 与 llms.txt；官网 Cloud 是单独候补，本条只写开源路径）；figma-specs（Figma 2026-10-15 博客，画布上的多人 Markdown，编程助手经 Figma MCP 读；开放测试。博客写付费方案，帮助中心写 Professional / Education / Organization / Enterprise 且需编辑权）。Agent UI：figma-pilot（MIT，README 工具表只有 figma_status、figma_execute、figma_get_api_docs，npm @youware-labs/figma-pilot-mcp）。三条均为 3:2 WebP。本库未装桌面版或插件、未接 MCP、未在付费工作区打开 Specs。cases 173→175（1.4.57→1.4.58），agent-ui 64→65（1.4.28→1.4.29）。未改首页精选、HTML、Models。未合并仍打开的 #104（Awesome DESIGN.md / photocn / Open Components / Blobz）与 #100（Impeccable / XIAS / SolidCraft）。AI×design: openpencil (MIT; local read/write of .fig/.pen; CLI @open-pencil/cli, MCP @open-pencil/mcp, Vue SDK and llms.txt; hosted Cloud is a separate waitlist, so this entry covers the open-source path only); figma-specs (Figma blog 2026-10-15; multiplayer Markdown on the canvas that coding agents read through Figma MCP; open beta. The blog says paid plans; the help center lists Professional / Education / Organization / Enterprise and requires edit access). Agent UI: figma-pilot (MIT; README tool table is only figma_status, figma_execute, and figma_get_api_docs; npm @youware-labs/figma-pilot-mcp). All three are 3:2 WebP. This library did not install the desktop app or plugin, did not connect MCP, and did not open Specs in a paid workspace. cases 173→175 (1.4.57→1.4.58), agent-ui 64→65 (1.4.28→1.4.29). Home picks, HTML, and Models unchanged. Did not merge still-open #104 (Awesome DESIGN.md / photocn / Open Components / Blobz) or #100 (Impeccable / XIAS / SolidCraft).
AI×设计优先，补今早之后出现的公开源：rayden-ui（MIT React/Tailwind + MCP）；affectui（AI 产品向 React 组件，仓库 MIT，条目只标免费路径）；geneva（当日 Show HN：JSON 时间线 + HTML/CSS 叠层的 MIT 命令行视频工具）。三条均为 Playwright 3:2 WebP。html-items 135→136，agent-ui 63→64，cases 172→173。未改首页精选、Models。未合并仍打开的晨更 PR #100（Impeccable / XIAS / SolidCraft）。AI×design-first catch-up after morning: rayden-ui (MIT React/Tailwind + MCP); affectui (AI-product React components, MIT repo, free path only); geneva (same-day Show HN: MIT CLI with JSON timelines and HTML/CSS overlays). All three use Playwright 3:2 WebP. html-items 135→136, agent-ui 63→64, cases 172→173. Home picks and Models unchanged. Did not merge still-open morning PR #100 (Impeccable / XIAS / SolidCraft).
整站梳理：分类、收录理由与页面一致性Site cleanup: categories, listing notes, and page consistency
Cases 新增 kind 字段（工具 / 教程 / 用法配方 / 资讯 / 实践案例），目录按它自动归档，标题去掉「工具：」等前缀；7 个放错到「学习」的新工具回到「工具与应用」，资讯不再混进教程。收录理由与摘要里的 HTTP 状态、star、HN 编号、截图方式移到详情页折叠的「核验记录」，构建会检查。精选按统一规则收紧到约 16%，难度 beginner 并入 starter。旧列表页标 noindex 并移出 sitemap，详情面包屑统一为「首页 / 目录 / 标题」，「品牌收集」移入「更多」。首页首屏与创意轨重排，模型页改用统一目录版式，补齐浅色模式的文字对比度。条目数量不变。Cases gain a kind field (tool / guide / playbook / news / case study); directories derive from it and titles drop type prefixes. Seven recent tools filed under Learning move to Tools, and launch news no longer mixes with guides. HTTP status, star counts, HN ids, and capture notes move out of listing notes and summaries into a folded Verification log on detail pages, with a build check. Featured follows one rule and drops to about 16%; beginner merges into starter. Legacy listings are noindex and leave the sitemap; breadcrumbs read Home / section / title; Brand collection moves into More. The homepage lead and creative rail are rebalanced, the Models page uses the shared directory layout, and light-mode text contrast is fixed. Entry counts unchanged.
收录：Cases +1（小王子 3D 剧情游戏）Catalog: Cases +1 (The Little Prince 3D story game)
Hao 指定收录 @Saccc_c 帖：16 天用 Opus 5.5 + Tripo + Blender 做的《小王子》3D 剧情网页游戏（littleprince.dev）。标题页可打开；无头浏览器无 WebGPU，未实玩；开源为作者计划，暂无仓库。Playwright 3:2 WebP 预览。cases 171→172。未改首页精选、Models。Hao-requested: @Saccc_c's Little Prince 3D story web game built in 16 days with Opus 5.5 + Tripo + Blender (littleprince.dev). Title page loads; not played (no WebGPU headless); open-source is planned, no repo yet. Playwright 3:2 WebP preview. cases 171→172. Homepage picks and Models unchanged.
AI×设计优先：suta（给 Claude Code/Codex 的界面 Skill，压 AI 同质化；Show HN https://news.ycombinator.com/item?id=50040224）；ns-ui（542 组件登记处，源码安装 + MCP/CLI；Show HN https://news.ycombinator.com/item?id=50011077）；editui（Chrome 扩展点选界面批量改完贴给助手，MIT）。三条均为 Playwright 3:2 WebP 预览；本库未装插件/扩展、未接 MCP 实测。html-items 133→134，agent-ui 61→62，cases 169→170。未改首页精选、Models。AI×design first: suta (UI skill for Claude Code/Codex against AI-slop; Show HN https://news.ycombinator.com/item?id=50040224); ns-ui (542-component registry with source install + MCP/CLI; Show HN https://news.ycombinator.com/item?id=50011077); editui (Chrome extension to batch UI notes for coding agents, MIT). All three have Playwright 3:2 WebP previews; this library did not install plugins/extensions or exercise MCP. html-items 133→134, agent-ui 61→62, cases 169→170. Homepage picks and Models unchanged.
Cases（AI×设计优先）：diffui（扩散模型一次出多版界面再交助手落地；Show HN https://news.ycombinator.com/item?id=48995754）；css-studio（在真实网站上改样式，经 MCP 交给现有编程助手写回；Show HN https://news.ycombinator.com/item?id=47702196）；wunderui（Figma+React 同名设计系统，免费 MCP / DESIGN.md / llms.txt 给助手读）。Agent UI：moodi（MIT 小 3D 状态角色；Show HN https://news.ycombinator.com/item?id=50047937）。Models：thx-01（HAL-X 322M 开源决策模型，HF 博文 2026-10-10）。各条均为 Playwright 实拍 3:2 WebP 预览；本库未登录付费档、未复跑作者基准。Cases (AI×design first): diffui (diffusion multi-option UI then agent handoff; Show HN https://news.ycombinator.com/item?id=48995754); css-studio (edit the live site, MCP to your coding agent; Show HN https://news.ycombinator.com/item?id=47702196); wunderui (matching Figma+React design system with free MCP / DESIGN.md / llms.txt). Agent UI: moodi (MIT tiny 3D state character; Show HN https://news.ycombinator.com/item?id=50047937). Models: thx-01 (HAL-X 322M open decision model; HF blog 2026-10-10). Each has a Playwright 3:2 WebP preview; this library did not use paid tiers or re-run author benchmarks.
HTML：oil-ui（给 AI Agent 的界面设计 Skill，先出几版风格不同的小样再细化，MIT；线索 @I_am_oil_oil 10-11 帖）；m3e-canvas（浏览器里拖 Material 3 Expressive 界面草图并生成给 AI 编程工具的提示词，MIT；线索 @lnkiai 10-11 帖）；prompt-motion（200 多条 Claude 动效视频与对应提示词合集，10-10 新增自定义提示词）。Agent UI：agent-standup（把 Claude Code 子 Agent 画成带配音的视频站会，MIT，HN 50039630）。四条都有 3:2 自托管预览，均未完整实测。html-items 130→133，agent-ui 59→60。未改首页精选、Cases、Models。HTML: oil-ui (a UI design skill for AI agents that mocks up distinct directions before refining, MIT; lead: @I_am_oil_oil post on 10-11); m3e-canvas (sketch Material 3 Expressive screens in the browser and generate a prompt for AI coding tools, MIT; lead: @lnkiai post on 10-11); prompt-motion (200+ Claude-made motion videos with their prompts, customizable prompts added 10-10). Agent UI: agent-standup (Claude Code sub-agents drawn as a voiced video stand-up, MIT, HN 50039630). All four have self-hosted 3:2 previews; none fully tested hands-on. html-items 130→133, agent-ui 59→60. Homepage picks, Cases, and Models unchanged.
Cases：claude-dashboards-motion（Claude 官方文章 2026-10-08，动画讲解短片 + 实时数据看板 beta，AI×设计，已进 caseTools）。HTML：quake-srp（Show HN item 50016312，纯安全 Rust 移植的《雷神之锤》浏览器版，Claude Code 智能体团队完成）。另更新 claude-design：官方宣布 Design / Docs / Slides 去掉 beta、含免费档。新条目有 3:2 自托管预览，均未实测。cases 165→166，html-items 127→128。未改首页精选、Agent UI、Models。Cases: claude-dashboards-motion (Claude's official 2026-10-08 article; animated explainers plus live data dashboards in beta; AI×design; added to caseTools). HTML: quake-srp (Show HN item 50016312; Quake ported to safe Rust, playable in the browser, built by Claude Code agent fleets). Also updated claude-design: Design / Docs / Slides are out of beta on every plan including Free. New entries have self-hosted 3:2 previews; nothing tested hands-on. cases 165→166, html-items 127→128. Homepage picks, Agent UI, and Models unchanged.
Cases：claude-design（Anthropic Labs AI 设计工具，产品页 + 2026-09 设计系统更新，AI×设计）、flowstep（给编程助手的设计画布 + MCP，AI×设计）、grok-bot-shopify（官方 @bot https://x.com/bot/status/2108228871938359347 ，连 Shopify 查单/库存/商品）。HTML：svg-spark（HN item 50013931，10 个本地浏览器设计/开发小工具）。全部新条目有 3:2 自托管预览；三条 Case 已进 caseTools。均未实测。cases 162→165，html-items 126→127。未改首页精选、Agent UI、Models。Cases: claude-design (Anthropic Labs AI design tool; product page + September 2026 design-system update; AI×design), flowstep (design canvas + MCP for coding agents; AI×design), grok-bot-shopify (official @bot https://x.com/bot/status/2108228871938359347 — Shopify for orders/inventory/listings). HTML: svg-spark (HN item 50013931, 10 local browser design/dev utilities). Every new entry has a self-hosted 3:2 preview; three Cases added to caseTools. Nothing tested hands-on. cases 162→165, html-items 126→127. Homepage picks, Agent UI, and Models unchanged.
10月8日 · 星期四 · 2 条10/8 · Thursday · 2 items
晚间收录：Cases +2（Krea Photoshop 插件 / Framer 技能做设计规范）、HTML +2（时区文件数据故事 / demoscene-recomp）Evening catalog: Cases +2 (Krea Photoshop plugin / Framer Skills as a design system), HTML +2 (time zone data story / demoscene-recomp)
Cases：krea-photoshop-plugin（@krea_ai 10-07 发布，插件页 krea.ai/plugins/photoshop，AI×设计）、framer-skills-design-system（Framer Skills 9-22 上线，10-07 官方玩法帖，把设计规范写成技能让助手做新页面，AI×设计）。HTML：holding-up-the-internet（HN item 50002494，滚动数据故事）、demoscene-recomp（HN item 50002426，DOS 演示片重编译为 WebAssembly）。全部新条目都有 3:2 自托管预览。均未实测。cases 160→162，html-items 124→126。未改首页精选、Agent UI、Models。Cases: krea-photoshop-plugin (@krea_ai launch on 10-07, plugin page krea.ai/plugins/photoshop, AI×design) and framer-skills-design-system (Framer Skills launched 9-22; official how-to post 10-07 on writing your design system as a Skill so the agent builds new pages, AI×design). HTML: holding-up-the-internet (HN item 50002494, a scrolling data story) and demoscene-recomp (HN item 50002426, DOS demos recompiled to WebAssembly). Every new entry has a self-hosted 3:2 preview. Nothing tested hands-on. cases 160→162, html-items 124→126. Homepage picks, Agent UI, and Models unchanged.
早间收录：Cases +3（X 内置读取 / 幻灯片 / AI×设计对照）、HTML +3（ascii.rest / 艺术史博物馆 / BIGWORDS）、Models +2（GPT-6 Intelligent UI / Claude Haiku 5.5）Morning catalog: Cases +3 (built-in X reading / slide decks / AI×design comparison), HTML +3 (ascii.rest / art-history museum / BIGWORDS), Models +2 (GPT-6 Intelligent UI / Claude Haiku 5.5)
Cases：grok-bot-x-built-in 来自官方 @bot 推文 https://x.com/bot/status/2107949161878606089 （10-08 05:40 北京时间），Grok Bot 不连 X 也能搜、读、盯公开帖子；grok-bot-slide-decks 来自官方更新日志 v0.68.1（10-07），Bot 先给样张再出 PPT 或 Google Slides；quesma-opus-5-5-design 是 AI×设计对照文，同一句提示词让 Opus 5.5 和 GPT-6 Astra 各做一个 three.js 可视化，线索 HN item 49991903。HTML：ascii-rest（HN item 49993857，MIT 动态字符画组件，偏设计）、walkable-art-history（Show HN item 49992057）、bigwords-page（Show HN item 49994443）。Models：gpt-6-intelligent-ui（HN item 49996425，回答直接生成可交互界面，属 AI×设计）、claude-haiku-5-5（HN item 49996437）。全部新条目都有 3:2 自托管预览。均未实测，数字为官方或作者自述。cases 157→160，html-items 121→124，models 25→27。未改首页精选。Cases: grok-bot-x-built-in from the official @bot post https://x.com/bot/status/2107949161878606089 (05:40 on 10-08 Beijing time): Grok Bot can search, read, and monitor public X posts without the X connector; grok-bot-slide-decks from the official changelog v0.68.1 (10-07): Bots show sample slides, then deliver PowerPoint or Google Slides; quesma-opus-5-5-design, an AI×design comparison where one prompt had Opus 5.5 and GPT-6 Astra each build a three.js visualization, lead HN item 49991903. HTML: ascii-rest (HN item 49993857, MIT animated ASCII components, design-leaning), walkable-art-history (Show HN item 49992057), bigwords-page (Show HN item 49994443). Models: gpt-6-intelligent-ui (HN item 49996425; answers generate interactive UI, AI×design) and claude-haiku-5-5 (HN item 49996437). Every new entry has a self-hosted 3:2 preview. Nothing tested hands-on; figures are vendor or author claims. cases 157→160, html-items 121→124, models 25→27. Homepage featured picks unchanged.
Cases：收 graphical —— Josh Puckett 与 Adam Michela 的视觉主题工具，调颜色、字体、间距、图标、动效和 Base UI 组件状态，下载或用 shadcn 安装，附 GUI.md 和三个编程助手技能。首页 HTTP 200。购买按钮写 $49，读屏文本 previously $99。未购买、未下载。HN item 49971356，撰写时 34pts。收 protopie-mcp —— 官方博客 2026-10-07，Studio MCP 在 Pro，Code MCP 与 Dev View 在 Enterprise，Code MCP 列表加入 React Native。未登录、未连接。收 figma-agent-ga —— 2026-10-06 发布说明写设计助手正式版：库的 markdown 指南、看同事助手、搜文件、生成式插件、Starter 上的自定义技能，并开始计 AI 额度。帮助中心写所有计划、按席位决定能否改文件；同日博客写 Starter 有限、部分 Education。两套官方说法都保留。未登录。与库内 figma-bro 不是同一条。收 gamma-5 —— 只依据 Gamma 帮助中心：对话式新建、AI 编辑重做、PPT/PDF 保留版式、主题跟文件走、形状可自由拖放。未找到 gamma.app 上的 Gamma 5 发布博文，不用新闻稿里的用户数。未登录、未导入。收 openai-math-progress —— 公告页 curl 403，正文经阅读代理核对；仓库 openai/math 写 722 篇、372 族，Apache-2.0，并非都有 Lean。平均约 3 小时 ChatGPT Pro 为官方估算。模型未发布，所以不进模型板。未读论文、未跑 Lean。HN item 49984923，撰写时 981pts。Models：收 strands-decider-2b —— 约 19 亿参数的开源决策模型，Apache-2.0。官博 2026-10-01，仓库现以 v21 为参考、v19 仍在。分数是仓库自测，未复跑。与 clef、gutsy、jeeves、typesafe-jev、laya-system-one、ollaya 互链。HN item 49987076，撰写时 208pts。收 embeddinggemma-2 —— 模型卡写 7.4 亿参数，文字、图像、音频进同一嵌入空间，许可页为 Apache 2.0。表内分数未复跑。官博里的手机内存数字未抄入要点。HN item 49980487，撰写时 352pts。五条 Cases 与两条 Models 均有 3:2 自托管预览。cases 152→157（1.4.47→1.4.48）；models 23→25（1.1.18→1.1.19）。未改 html-items / agent-ui，未改首页精选。跳过：C64 Keyboard 字体仓库没有 SPDX 许可；Treg 是按次计费的工具目录，许可字段为空，不是这次要收的设计工具。未并入仍开放的 #87（google-stitch、brilliant-design、magic-patterns、figma-use、tegaki、onlook）。Cases: graphical — Josh Puckett and Adam Michela’s visual theme tool for colors, type, spacing, icons, motion, and Base UI component states, via download or shadcn, with GUI.md and three coding-agent skills. Homepage HTTP 200. The buy button says $49, with screen-reader text previously $99. Not purchased, not downloaded. HN item 49971356, 34pts at write time. protopie-mcp — official blog 2026-10-07: Studio MCP on Pro, Code MCP and Dev View on Enterprise, React Native added to the Code MCP list. Not signed in, not connected. figma-agent-ga — the 2026-10-06 release notes mark the design agent generally available: markdown library guidelines, watching teammates’ agents, file search, generative plugins, custom skills on Starter, and AI credits from that day. The help center says all plans, with the seat deciding who can edit; the same day’s blog says Starter is limited and only some Education plans qualify. Both official wordings are kept. Not signed in. Not the same entry as figma-bro. gamma-5 — help center only: conversation-first creation, rebuilt AI editing, PPT/PDF layouts kept, themes per file, shapes as freeform blocks. No Gamma 5 launch post found on gamma.app, so press user-counts are not used. Not signed in, nothing imported. openai-math-progress — the announcement curl returned 403 and the body was checked through a reader proxy; repo openai/math says 722 manuscripts and 372 families, Apache-2.0, not all in Lean. About three hours of ChatGPT Pro is the vendor estimate. The model is unreleased, so this is not a model card. Papers not read, Lean not run. HN item 49984923, 981pts at write time. Models: strands-decider-2b — an open decision model of about 1.9 billion parameters, Apache-2.0. Blog dated 2026-10-01; the repo now points at v21 and still publishes v19. Scores are the repo’s own tests, not re-run. Cross-linked with clef, gutsy, jeeves, typesafe-jev, laya-system-one, and ollaya. HN item 49987076, 208pts at write time. embeddinggemma-2 — the model card says 740 million parameters and one embedding space for text, images, and audio; the license page is Apache 2.0. Table scores not re-run. Phone-memory figures from the blog are not copied into the highlights. HN item 49980487, 352pts at write time. All five Cases and both Models have self-hosted 3:2 previews. cases 152→157 (1.4.47→1.4.48); models 23→25 (1.1.18→1.1.19). html-items / agent-ui and home picks unchanged. Skipped: the C64 Keyboard font repo has no SPDX license; Treg is a per-call tool catalog with an empty license field, and it is not a design tool for this pass. The still-open #87 entries (google-stitch, brilliant-design, magic-patterns, figma-use, tegaki, onlook) were not duplicated.
Cases：新开 /cases/artcraft-crafting-apps/ —— ArtCraft 团队用 Rust 从零写的七款开源创作软件，对标 Photoshop / Illustrator / Premiere Pro / Lightroom / Acrobat / After Effects / InDesign，每款都带 CLI 和 MCP 服务器，AI 助手可直接操作。线索为 X 帖 @math_fournity（北京时间 10-07 03:30）；以官网总览页、PhotoCraft 产品页和七个 storytold 仓库 README 为准（撰写时均 HTTP 200，MIT/Apache-2.0 双许可，声明与 Adobe 无关联）。写清是什么、适合谁、怎么试（下载/源码/MCP 接入）和边界：早期 alpha、官方自认替代不了日常专业工作、PSD 重新保存非逐字节相同；未下载、未安装、未接 MCP。预览为官网四款应用官方截图 3:2 拼图 WebP。同步加入 directory.json caseTools。cases 151→152（→1.4.47）。未改首页精选。Cases: new /cases/artcraft-crafting-apps/ — seven open-source creative apps the ArtCraft team wrote from scratch in Rust, matching Photoshop / Illustrator / Premiere Pro / Lightroom / Acrobat / After Effects / InDesign, each with a CLI and an MCP server so an AI agent can drive it. Lead was an X post by @math_fournity (03:30 on 10-07 Beijing time); the entry follows the official overview page, the PhotoCraft product page, and the seven storytold repo READMEs (all HTTP 200 at write time; dual MIT/Apache-2.0; not affiliated with Adobe). Covers what it is, who it is for, how to try it (installers / source / MCP), and the limits: early alpha, the team’s own note that it cannot yet replace daily professional work, and PSD re-saves not being byte-identical; nothing downloaded, installed, or connected over MCP. Preview is a 3:2 WebP collage of four official app screenshots. Added to directory.json caseTools. cases 151→152 (→1.4.47). Homepage featured picks unchanged.
Models：新开 /models/nano-banana-2-1/ —— Google 2026-10-06 发布的出图改图模型（X 帖 @Google 北京时间 10-07 00:00），主打版式设计、遮罩改图和主体一致。官方来源：X 帖、Gemini API 模型页、DeepMind 模型卡、出图指南、价格页、更新日志（撰写时均 HTTP 200）。写清是什么、适合谁、怎么试（Gemini App / AI Studio / API，API 无免费档）、模型卡自述的不足；对比分与价格标为官方自述，未出图、未复跑。预览为官方宣传短片 2 倍速 3:2 动图 WebP。同板关联 gemini-4-argon。models 22→23（→1.1.18）。未改首页精选。Models: new /models/nano-banana-2-1/ — Google’s image generation and editing model released 2026-10-06 (X post by @Google at 00:00 on 10-07 Beijing time), focused on layout design, masked editing, and subject consistency. Official sources: the X post, Gemini API model page, DeepMind model card, image generation guide, pricing page, and release notes (all HTTP 200 at write time). Covers what it is, who it is for, how to try it (Gemini app / AI Studio / API; the API has no free tier), and the limits the model card itself lists; scores and prices marked as vendor claims, no images generated, nothing re-run. Preview is the official promo clip at 2× speed as a 3:2 animated WebP. Links to gemini-4-argon on the board. models 22→23 (→1.1.18). Homepage picks unchanged.
Cases：收 james-pember-grok-bot-team —— James Pember 的 Substack 实测（2026-10-04，HN item 49960401），两个可照做的用法：Bot 在 Slack 里先分诊产品问题，以及双周会前自动汇总发布说明、GitHub 和 Slack 做会前材料；未复跑。HTML：收 banyan（Show HN 49977811，单 HTML 文件的程序化榕树，AGPL-3.0）、gss-lang（Show HN 49977287，用 CSS 写 3D 场景、编译成一个着色器，Apache-2.0）、frutjam（Show HN 49978580，纯 CSS 的 Tailwind v4 组件库，MIT）。Agent UI：收 shaders-com（HN 49980593，MIT 开源 WebGPU 特效组件，免费档含 MCP；Pro 预设、区块与提示词不收）。五条均配 3:2 自托管预览。官方 @bot 与 x.ai changelog 自 10-02 v0.66.0 后无新条目，本次不加。cases 150→151（→1.4.46）；html-items 118→121（→1.3.56）；agent-ui 57→58（→1.4.22）。未改首页精选。Cases: james-pember-grok-bot-team — James Pember’s Substack write-up (2026-10-04, HN item 49960401) with two copyable workflows: a Bot that triages product questions in Slack first, and a Bot that compiles release notes, GitHub, and Slack into a pre-read before the fortnightly sync; not re-run. HTML: banyan (Show HN 49977811, a procedural banyan in one HTML file, AGPL-3.0), gss-lang (Show HN 49977287, CSS-style 3D scenes compiled to one shader, Apache-2.0), frutjam (Show HN 49978580, CSS-only Tailwind v4 components, MIT). Agent UI: shaders-com (HN 49980593, MIT open-source WebGPU effect components with a free MCP; Pro presets, sections, and prompts excluded). All five have self-hosted 3:2 previews. No new official @bot or x.ai changelog entries since v0.66.0 on 10-02. cases 150→151 (→1.4.46); html-items 118→121 (→1.3.56); agent-ui 57→58 (→1.4.22). Homepage featured picks unchanged.
10月6日 · 星期二 · 2 条10/6 · Tuesday · 2 items
晚间收录：Models +1（Beam）、Cases +1（CoIsland）、HTML +3（Samon / Thank you, Steve / Pumpkin Patch）Evening catalog: Models +1 (Beam), Cases +1 (CoIsland), HTML +3 (Samon / Thank you, Steve / Pumpkin Patch)
Models：新开 /models/beam/ —— Reflection 2026-10-05 官博的开源权重稀疏混合专家，总参数 5010 亿、每次启用 230 亿。官博与 platform.reflection.ai 撰写时 HTTP 200。博文写仍在最终安全测试，权重、技术报告、模型卡和开发者配套这个月晚些时候才放，许可打算 Apache 2.0；现在不能下载，本库未登记。表格分数、3–4 倍推理计算量、23.8 万亿 token、GPU 与尝试次数均为厂商自述，未复跑。Artificial Analysis 的 /models/beam 与 /models/reflection-beam 均为 HTTP 404，不编第三方榜。HN item 49969183，撰写时 HN 页面与 Algolia 都是 478pts，2026-10-05T19:16:35.000Z。精选，不打星。与 gemini-4-argon、gpt-6-astra、claude-opus-5-5 互链。Cases：收 coisland —— Mac 刘海应用，把 GitHub / Vercel / Sentry / Jira / Snowflake 等告警一键交给 Claude Code、Codex 或 Cortex Code。首页 Chrome HTTP 200（本环境 curl 被 Cloudflare 拦成 403）；连接器、助手、隐私页可读文本 200。Show HN item 49970043，撰写时 9pts。免费为最早 2 条监控；早鸟 $29.99 至 10 月 25 日，之后 $49.99，一次买断、最多 2 台 Mac。苹果自研芯片、macOS 14+，不支持 Intel。隐私政策比「没有服务器」更细：还有更新服务器和买后的许可证服务器，不收工具数据。未安装、未把告警交给助手。配富落地页与 3:2 预览。HTML：收 samon（HN 49972211，47pts）gwtar 设计文档，可玩原型在 #play，148 座花园，求解器数字未重跑，CC0；收 thank-you-steve（Show HN 49964102，15pts）Akash Bangad 的滚动纪念，预览是开场点阵肖像，没有滚完后段；收 pumpkin-patch（Show HN 49965227，24pts）一百万只南瓜，截图时已刻 447 只，VR 入口 HTTP 200，未认领、未进头显。三条均有 3:2 自托管预览。Agent UI 无新增。models 21→22（1.1.16→1.1.17）；cases 149→150（1.4.44→1.4.45）；html-items 115→118（1.3.54→1.3.55）。未改首页精选。Models: new card /models/beam/ — Reflection’s 2026-10-05 open-weight sparse mixture-of-experts, 501 billion parameters total and 23 billion active. The blog and platform.reflection.ai returned HTTP 200. The post says final safety testing is still underway, and the weights, technical report, model card, and developer extras come later this month under a planned Apache 2.0 license. Not downloadable; this library did not sign up. Table scores, the 3–4× inference-compute claim, 23.8 trillion tokens, and the GPU and rollout counts are vendor claims, not re-run. Artificial Analysis /models/beam and /models/reflection-beam both returned HTTP 404, so no third-party board link. HN item 49969183, 478pts on both the HN page and Algolia at write time, 2026-10-05T19:16:35.000Z. Featured, unrated. Cross-linked with gemini-4-argon, gpt-6-astra, and claude-opus-5-5. Cases: coisland — a Mac notch app that hands alerts from GitHub / Vercel / Sentry / Jira / Snowflake and the rest to Claude Code, Codex, or Cortex Code in one click. Homepage HTTP 200 in Chrome (curl from this environment is Cloudflare 403). Connectors, agents, and privacy pages returned 200 as readable text. Show HN item 49970043, 9pts at write time. Free tier is the two oldest monitors. Early bird $29.99 until October 25, then $49.99, paid once, up to 2 Macs. Apple silicon, macOS 14+, not Intel. The privacy policy is more specific than “no server”: an update server and, after purchase, a license server exist and do not receive tool data. Not installed, no alert handed over. Rich landing and a 3:2 preview. HTML: samon (HN 49972211, 47pts), a gwtar design doc with the playable prototype at #play, 148 gardens, solver figures not re-run, CC0; thank-you-steve (Show HN 49964102, 15pts), Akash Bangad’s scrolling tribute, preview is the opening dotted portrait, later scenes not scrolled; pumpkin-patch (Show HN 49965227, 24pts), a million pumpkins, 447 carved at capture time, VR URL HTTP 200, not claimed and not walked in a headset. All three have self-hosted 3:2 previews. No Agent UI additions. models 21→22 (1.1.16→1.1.17); cases 149→150 (1.4.44→1.4.45); html-items 115→118 (1.3.54→1.3.55). Homepage picks unchanged.
早间收录：Cases +7（官方 Guides 5 篇 + 采购 / 客服两篇官方实践）、HTML +3（Flatten SF / stillwet / Island Three）Morning catalog: Cases +7 (five official Guides plus procurement and support case studies), HTML +3 (Flatten SF / stillwet / Island Three)
Grok Bot 案例补上 9 月 24 日一批官方 Guides：for Work、Post-Sales、Recruiting、SDRs、Legal Ops（https://x.ai/bot/guides），以及官方 News 的 Haggle Bot 采购实践（https://x.ai/news/grok-bot-procurement）和客服规模化实践（https://x.ai/news/grok-bot-customer-support）。HTML 收集：Flatten SF（HN item 49971230）、stillwet AI 油画回放（Show HN 49928566）、Island Three 浏览器太空圆柱城（Show HN 49963508），均带 3:2 自托管预览。@bot 今早没有新的可复现用法帖；Agent UI 今天没有新增。Cases adds the Sep 24 batch of official Guides: for Work, Post-Sales, Recruiting, SDRs, and Legal Ops (https://x.ai/bot/guides), plus the official Haggle Bot procurement post (https://x.ai/news/grok-bot-procurement) and the support-at-scale post (https://x.ai/news/grok-bot-customer-support). HTML Collection: Flatten SF (HN item 49971230), stillwet AI painting replays (Show HN 49928566), and Island Three, a walkable O’Neill cylinder (Show HN 49963508), each with a self-hosted 3:2 preview. No new reproducible @bot usage posts this morning; no Agent UI additions today.
10月5日 · 星期一 · 3 条10/5 · Monday · 3 items
晚间收录：Cases +2（SCM Screen Memories / Cloudflare Web Search API）Evening catalog: Cases +2 (SCM Screen Memories / Cloudflare Web Search API)
Cases：收 scm-screen-memories —— 本机 macOS 找图应用 SCM（Screen Memories）。仓库 https://github.com/allenv0/SCM 撰写时 gh api MIT、JavaScript、创建于 2026-10-03T16:11:55Z、329★、Releases 0；产品页 https://scm.allenlee.site HTTP 200；Show HN item 49952111（2026-10-04T09:24:52.000Z，154pts）。五种搜索（Files / Scenes / OCR / Dialogue / 可选 LLMs）与 Homebrew 安装（Apple 芯片、macOS 12+，默认 CLIP 权重大约 435MB）按 README 记录。未安装、未执行 brew。收 cloudflare-web-search-api —— Cloudflare 2026-10-02 changelog 的 Web Search API（beta）。页与用法文档 HTTP 200。供应商 Ceramic.ai / Exa / Linkup，经 AI Gateway 按标价计费、可自带密钥。HN item 49963171（2026-10-05T10:47:06.000Z，HN 页面 65pts / Algolia 61pts，撰写时仍在首页）。未调用。Cases 无 relatedModelIds，故在来源和试步骤链到已有 Clef 模型页（HTTP 200）与官博，不改 models.json。两条均配富落地页与 Playwright 3:2 预览。cases 140→142（1.4.42→1.4.43）。未改 html-items / models / agent-ui，未改首页精选。未动未合的 #78 / #79 / #83。Cases: added scm-screen-memories — SCM (Screen Memories), a local macOS photo-and-video finder. Repo https://github.com/allenv0/SCM at write time via gh api: MIT, JavaScript, created 2026-10-03T16:11:55Z, 329★, 0 releases; product page https://scm.allenlee.site HTTP 200; Show HN item 49952111 (2026-10-04T09:24:52.000Z, 154pts). Five modes (Files / Scenes / OCR / Dialogue / opt-in LLMs) and the Homebrew install (Apple silicon, macOS 12+, default CLIP weights about 435MB) are recorded from the README. Not installed, brew not run. Added cloudflare-web-search-api — Cloudflare’s Web Search API (beta) from the 2026-10-02 changelog. Page and how-to HTTP 200. Providers Ceramic.ai / Exa / Linkup, billed through AI Gateway at list price, bring-your-own key supported. HN item 49963171 (2026-10-05T10:47:06.000Z, HN page 65pts / Algolia 61pts, still on the front page at write time). Not called. Cases have no relatedModelIds, so the sources and try-steps link the existing Clef model page (HTTP 200) and the blog; models.json was not edited. Both ship with a rich landing and a Playwright 3:2 preview. cases 140→142 (1.4.42→1.4.43). html-items / models / agent-ui and home picks unchanged. Open PRs #78 / #79 / #83 were left untouched.
Cases +1：发帖图片素材站清单（@0xCzyan）Cases +1: posting image source sites list (@0xCzyan)
Cases：收 image-source-sites-0xczyan —— 子砚 @0xCzyan 2026-10-04 推文 https://x.com/0xCzyan/status/2106663001164247100 （约 16:29 CST）列出六类发帖素材站：Know Your Meme / Imgur / Giphy / Unsplash / Pexels / Wallhaven。fxtwitter 核对 200；五站 curl 200，Unsplash 本环境 401（反爬），未编通断。配富落地页与原帖配图 3:2 预览。不是 AI 工具，收成内容资源笔记；许可自负。cases 139→140（1.4.41→1.4.42）。未改 html-items / models / agent-ui。与未合的 #78 / #79 可能在 cases.json / changelog.json 冲突，本 PR 基于最新 main。Cases: added image-source-sites-0xczyan — Ziyan @0xCzyan’s 2026-10-04 post https://x.com/0xCzyan/status/2106663001164247100 (about 16:29 CST) listing six posting asset sites: Know Your Meme / Imgur / Giphy / Unsplash / Pexels / Wallhaven. fxtwitter check 200; five sites curl 200, Unsplash 401 here (bot filter), reachability not invented. Ships with a rich landing and a 3:2 crop of the post image. Not an AI tool — filed as a content-resource tip; licenses are on you. cases 139→140 (1.4.41→1.4.42). html-items / models / agent-ui unchanged. May conflict with unmerged #78 / #79 on cases.json / changelog.json; this PR is based on latest main.
早间收录：Cases +1（Main Bot 主动建议）、HTML +3（VB6 Studio Web / 全球灯塔 / 遥远未来时间线）Morning catalog: Cases +1 (Main Bot proactive suggestions), HTML +3 (VB6 Studio Web / Lighthouses of the World / Timeline of the Far Future)
Cases：收 main-bot-proactive —— 官方 @bot 推文 https://x.com/bot/status/2105713240701538538 （2026-10-01，主 Bot 主动提出能帮的事，建议不计用量）与 changelog v0.66.0 https://x.ai/changelog/bot （可设带星标的 Main Bot，主动跟进并协调其他 Bot，需 0.59.0+）。未实测。HTML：收 vb6-studio-web（HN item 49956681，69pts，MIT）；lighthouses-of-the-world（HN item 49933461，161pts，75,919 盏灯按真实节奏闪）；timeline-far-future（Show HN item 49940317，18pts，three.js 改编维基百科，CC BY-SA 4.0）。三条均有 Playwright 实拍 3:2 自托管预览。Agent UI 无新增：ThreeUI Community 仍为 npm 1.2.0（2026-09-01），仓库无新组件。Cases: added main-bot-proactive — official @bot post https://x.com/bot/status/2105713240701538538 (2026-10-01; the primary Bot offers help unprompted, suggestions don’t count against usage) and changelog v0.66.0 https://x.ai/changelog/bot (choose a starred Main Bot that checks in and coordinates other Bots; needs 0.59.0+). Not tested. HTML: added vb6-studio-web (HN item 49956681, 69pts, MIT); lighthouses-of-the-world (HN item 49933461, 161pts, 75,919 lights blinking on real cycles); timeline-far-future (Show HN item 49940317, 18pts, three.js take on the Wikipedia page, CC BY-SA 4.0). All three have self-hosted Playwright 3:2 previews. Agent UI: nothing new; ThreeUI Community is still npm 1.2.0 (2026-09-01) with no new components.
最新模型：新开 /models/gutsy/ —— kouhxp 的 gutsy-0.8b v0.4，本机 CPU 可跑的 Jev 兼容决策模型（不生成自由文本）。仓库 https://github.com/kouhxp/gutsy 与 HF https://huggingface.co/kouhxp/gutsy 晚间 HTTP 200；gh api Apache-2.0、20★；HF likes 2 / downloads 69。Show HN item 49923223，Algolia 晚间核对 9pts，2026-10-01T15:44:42Z。JevBench / 面板分与延迟标作者自测，未复跑、未代下 GGUF。与 typesafe-jev、laya-system-one、clef、jeeves、ollaya 互链；relatedCaseIds 指向 jeff 与 onesie。Cases：收 television —— Telepath 刚公开的个人 Agent GUI（artifact 工作台）。产品页 https://television.run/、仓库 telepath-computer/television、Show HN item 49939817（2026-10-02T23:25:04Z / 约 2026-10-03 07:25 CST，Algolia 晚间 6pts）晚间均 HTTP 200；README 写 MIT（GitHub license 字段 NOASSERTION）。配富落地页与 Playwright 3:2 预览。跳过：HN newest 上过薄的 Screen Loupe / 非 AI 条目；AgentComposerUI（10-01）信号一般且偏组件库、今晚未配预览；ShowKit / Demoday 与已有 HTML demo 工具重叠且未核到今日强信号；Karumi 付费 agentic demo 不符 Agent UI 免费开源口径；X API 受阻，未发帖。models 20→21（1.1.15→1.1.16）；cases 137→138（1.4.39→1.4.40）。未改 html-items / agent-ui，未改首页精选。Models: new card /models/gutsy/ — kouhxp’s gutsy-0.8b v0.4, a Jev-compatible decision model that runs on ordinary CPU and does not generate free text. Repo https://github.com/kouhxp/gutsy and HF https://huggingface.co/kouhxp/gutsy returned HTTP 200 this evening; gh api Apache-2.0, 20★; HF likes 2 / downloads 69. Show HN item 49923223, 9pts on Algolia at the evening check, 2026-10-01T15:44:42Z. JevBench / panel scores and latency are author self-tests — not re-run, GGUF not downloaded. Cross-linked with typesafe-jev, laya-system-one, clef, jeeves, ollaya; relatedCaseIds point at jeff and onesie. Cases: television — Telepath’s just-opened GUI for personal agents (artifact workspace). Product page https://television.run/, repo telepath-computer/television, and Show HN item 49939817 (2026-10-02T23:25:04Z / about 2026-10-03 07:25 CST, 6pts on Algolia this evening) all HTTP 200; README says MIT (GitHub license field NOASSERTION). Ships with a rich landing block and a Playwright 3:2 preview. Skipped: thin HN newest items such as Screen Loupe / non-AI posts; AgentComposerUI (Oct 1) only moderate signal and no preview tonight; ShowKit / Demoday overlap existing HTML-demo tooling without a strong same-day spike; Karumi is a paid agentic demo and fails the free/open Agent UI bar; X API blocked, no post. models 20→21 (1.1.15→1.1.16); cases 137→138 (1.4.39→1.4.40). html-items / agent-ui and home picks unchanged.
最新模型：新开 /models/clef/ —— Cloudflare 2026-10-01 发布的开源决策模型 Clef 与 Clef-Flash。官博 https://blog.cloudflare.com/clef-decision-models/ 、Workers AI changelog、HF Cloudflare/clef 与 clef-flash（Apache-2.0）晚间均为 HTTP 200。Workers AI：@cf/cloudflare/clef（27B）与 @cf/cloudflare/clef-flash（9B）；官方称 Jev / System One API 兼容，状态可为文本 / JSON / 图像 / 视频。HF 卡写骨干为 Qwen/Qwen3.8-27B 与 Qwen/Qwen3.5-9B（厂商 / HF 口径）。Decision Index、延迟与工作流表标厂商称，未复跑。RL 微调记为先 FDE、后自助的路线图。HN item 49923692，Algolia 晚间核对 536pts，2026-10-01T16:18:57Z。HF 晚间 likes 547 / 183。Artificial Analysis clef URL HTTP 404，不编第三方榜。晨间跳过非 A/B/C 头条，晚间补 Models。不打星。与 typesafe-jev、laya-system-one、jeeves、ollaya 互链并回写 relatedModelIds；relatedCaseIds 指向 jeff 与 onesie。Cases：收 deepseek-harness-desktop —— 官方产品页全球公开预览的 macOS / Windows 桌面端（另有 `npx @deepseek-ai/dsh web`），Cordis 插件架构；HN item 49929489，晚间 262pts；仓库 MIT，gh api 242104★。Experimental 以页面标签为准；未代装、未登录、不写第三方赠金。收 weave-router —— 晨间点名的晚间候选。Show HN item 49911500，105pts，无外链；仓库 weave-os/router Apache-2.0，5525★（workweave/router 301）。README：`npx @weave-os/router`，`--local` 自托管；Terminal-Bench / SWE-Atlas 相对 GPT-6 Astra 的数字标厂商自述。两条均配富落地页与 Playwright 3:2 预览。跳过：Audionaut（Show HN 49931031，Algolia 晚间 32pts）开源多轨音频编辑器，AI 目录契合弱；Janus GGUF Vulkan 早间已因过薄跳过；Frog and Toad 随笔不是工具；SvelteKit 3 是框架，不进目录；Microsoft MAI-Transcribe-2-Streaming 过新、信号弱；Agent UI 无足够确认的新免费可复制件；X API 受阻，未发帖。models 19→20（1.1.14→1.1.15）；cases 135→137（1.4.38→1.4.39）。未改 html-items / agent-ui，未改首页精选。Models: new card /models/clef/ — Cloudflare’s open-weight decision models Clef and Clef-Flash, announced 2026-10-01. The blog, Workers AI changelog, and HF Cloudflare/clef plus clef-flash (Apache-2.0) all returned HTTP 200 this evening. Workers AI: @cf/cloudflare/clef (27B) and @cf/cloudflare/clef-flash (9B); vendor claims Jev / System One API compatibility and text / JSON / image / video state. HF cards name the backbones Qwen/Qwen3.8-27B and Qwen/Qwen3.5-9B (vendor / HF claim). Decision Index, latency, and workflow tables are vendor claims, not re-run. RL fine-tuning is recorded as FDE first, self-serve later. HN item 49923692, 536pts on Algolia at the evening check, 2026-10-01T16:18:57Z. HF evening likes 547 / 183. The Artificial Analysis clef URL returned HTTP 404 — no third-party board link. Morning skipped non-A/B/C headlines; evening adds the Models card. Unrated. Cross-linked with typesafe-jev, laya-system-one, jeeves, and ollaya (relatedModelIds written back); relatedCaseIds point at jeff and onesie. Cases: deepseek-harness-desktop — official public-preview macOS / Windows desktop (also `npx @deepseek-ai/dsh web`) on a Cordis plugin architecture; HN item 49929489, 262pts; repo MIT, 242104★ on the gh api. Experimental follows the page label; not installed, not signed in, no third-party promo. weave-router — the evening candidate the morning note deferred. Show HN item 49911500, 105pts, no external URL; repo weave-os/router Apache-2.0, 5525★ (workweave/router 301). README: `npx @weave-os/router`, `--local` to self-host; Terminal-Bench / SWE-Atlas figures versus GPT-6 Astra are vendor claims. Both ship with a rich landing and a Playwright 3:2 preview. Skipped: Audionaut (Show HN 49931031, 32pts on Algolia this evening), an open multitrack audio editor and a weak AI-catalog fit; Janus GGUF Vulkan, already skipped as thin this morning; the Frog and Toad essay, not a tool; SvelteKit 3, a framework, not a catalog entry; Microsoft MAI-Transcribe-2-Streaming, too new and low-signal; Agent UI, no clear new free piece; X API blocked, nothing posted. models 19→20 (1.1.14→1.1.15); cases 135→137 (1.4.38→1.4.39). html-items / agent-ui unchanged; homepage picks unchanged.
早间收录：Cases +1（Pi 1.0）、HTML +2（CSS Bed / Wave）Morning catalog: Cases +1 (Pi 1.0), HTML +2 (CSS Bed / Wave)
Cases：收 pi-1-0 —— HN 首页 item 49926069（Algolia 早间核对 709pts，2026-10-01T19:33:05Z）https://earendil.com/posts/pi-1-0/ 。Earendil MIT 极简 agent harness 发 1.0（Codemode / 虚拟模型 / 延迟工具 / cache warming 等）；同日 Pi Durable（item 49925969，206pts）。仓库 earendil-works/pi 早间 111225★；pi.dev HTTP 200。配富落地页与 Playwright 3:2 预览。HTML：收 css-bed（HN item 49927212，54pts）https://www.cssbed.com/ —— classless 主题对照床，仓库约 27 套主题、183★、无 SPDX LICENSE；与已收 Pico CSS 相邻但不重复。收 wave-subworkflow（Show HN item 49843746，7pts）https://wave.subworkflow.ai/ —— Three.js 层叠波浪渐变编辑器，可即时拖拽试。两条均配自托管 3:2 预览。跳过：Agent UI（ThreeUI Community 自 2026-09-01 同步后无足够确认的新免费可复制件）；Weave Router（Show HN 49911500 / workweave/router，晚间候选）；Janus / Breadcrumb 等分低或未沉淀；@bot X API client-not-enrolled，未硬刷。cases 134→135（1.4.37→1.4.38）；html-items 107→109（1.3.51→1.3.52）。未改 agent-ui。Cases: pi-1-0 — HN front page item 49926069 (709pts morning Algolia, 2026-10-01T19:33:05Z) https://earendil.com/posts/pi-1-0/. Earendil MIT minimal agent harness 1.0 (Codemode / virtual models / deferred tools / cache warming, etc.); same-day Pi Durable (item 49925969, 206pts). Repo earendil-works/pi ~111225★ morning; pi.dev HTTP 200. Rich landing + Playwright 3:2 preview. HTML: css-bed (HN item 49927212, 54pts) https://www.cssbed.com/ — classless theme bed (~27 themes, 183★, no SPDX LICENSE); neighbor to Pico CSS, not a duplicate. wave-subworkflow (Show HN item 49843746, 7pts) https://wave.subworkflow.ai/ — tryable Three.js layered wave-gradient editor. Both with self-hosted 3:2 previews. Skipped: Agent UI (no clear new free ThreeUI Community pieces since 2026-09-01 sync); Weave Router (Show HN 49911500 / workweave/router, evening candidate); thin Janus / Breadcrumb; @bot X API client-not-enrolled, not hammered. cases 134→135; html-items 107→109; agent-ui unchanged.
最新模型：新开 /models/gpt-6-1-sol/ —— OpenAI 于 2026-09-29 发布的 GPT-6 Sol 升级款。官方页 https://openai.com/index/introducing-gpt-6-1-sol/（标题 “Introducing GPT-6.1 Sol”，页内 “Near-Astra intelligence for a fifth of the price”）晚间 Playwright GET HTTP 200。官方称在 agentic coding、电脑操作与专业工作上几乎对齐 GPT-6 Astra，标准输入与输出 token 价格约为 Astra 的五分之一。标准 API：$2 / 百万输入、$0.10 / 百万缓存输入、$10 / 百万输出；缓存输入官方称比标准输入少 95%、比 GPT-6 Sol 缓存输入少 50%。API id：gpt-6.1-sol。可用性：Plus / Pro / Business / Enterprise / Edu 的 ChatGPT Work 与 Codex；尚未进入 Chat；API 已可用。Ultrafast 官方称随后数日推出，在 Codex 中相对标准速度最高 8 倍 token 生成。官方评测摘录（标官方称，本库未复跑）：DeepSWE v1.1 约五分之一成本对齐 Astra，并以更低 effort/成本超出 GPT-6 Sol 最佳 6.4 个百分点；GDP.pdf 高于带 fallbacks 的 Opus 5.5、成本不到一半；AutomationBench 在 medium effort 比 Opus 5.5 高 2.2 个百分点、约三分之一成本，相对 GPT-6 Sol +4.8 个百分点（页内措辞）；OSWorld 2.0 离线 partial 在 max effort 比 GPT-6 Sol 高 7 个百分点、成本不到一半，距 Astra 2.1 个百分点以内；Terminal-Bench Science 0.1 在 max effort 超过 GPT-6 Sol 一倍以上、单任务成本不到一半，Astra 在最难科学任务上仍最高（68.1%）；难提示事实性低 effort 含事实错误占比相对 GPT-6 Sol 从 11.4% 降到 7.7%。第三方榜 Artificial Analysis https://artificialanalysis.ai/models/gpt-6-1-sol（GPT-6.1 Sol (max)，晚间 GET 200）：Intelligence Index 52（#10/222），67.8 tok/s（#107/222），$2/$10、Cache Discount 95%、每任务 $0.72，1M 上下文，发布日 2026-09-29。AA 文章 https://artificialanalysis.ai/articles/gpt-6-1-sol-replaces-gpt-6-sol-after-just-7-days-with-near-astra-intelligence 称 Sol 在 7 天后换成 6.1 Sol，约比 Astra II 低 1 分、每任务成本不到四分之一。HN item 49896586，Algolia 晚间核对约 974pts，2026-09-29T17:06:45Z / 约 2026-09-30 01:06 CST。系统卡附录 https://deploymentsafety.openai.com/gpt-6-1-sol 晚间 HTTP 200，未逐条核评测。晨间只收 Cases + HTML，并明确跳过通用 DevDay 新闻、不进 A/B/C，故 Models 当时没有 gpt-6-1-sol；晚间补上。精选，整理质量 4 星。与 gpt-6-astra、astra-openai-showcase、claude-sonnet-5-5 互链，并回写 gpt-6-astra 与 claude-sonnet-5-5 的 relatedModelIds。预览为 Playwright 实拍官方页首屏裁 3:2 WebP。跳过：薄 Show HN（约 1–3pts）TabX / Rhubarb / Instapath / Moching / Fab；HTML 的 Space / Enigma / Scissor 晨间已收；Agent UI 无足够确认免费的 Community 新件，未硬凑；Cases 无新的 Grok Bot 指南。models 17→18（1.1.12→1.1.13）。未改 cases / html-items / agent-ui。Models: new card /models/gpt-6-1-sol/ — OpenAI’s upgrade to GPT-6 Sol, announced 2026-09-29. Official page https://openai.com/index/introducing-gpt-6-1-sol/ (title “Introducing GPT-6.1 Sol”, on-page “Near-Astra intelligence for a fifth of the price”) returned HTTP 200 on an evening Playwright GET. Vendor claim: it nearly matches GPT-6 Astra on agentic coding, computer use, and professional work at about one-fifth of Astra’s standard input and output token prices. Standard API: $2 per million input, $0.10 per million cached input, $10 per million output; cached input claimed 95% less than standard input and 50% less than GPT-6 Sol cached input. API id gpt-6.1-sol. Availability: Plus / Pro / Business / Enterprise / Edu in ChatGPT Work and Codex; not yet in Chat; API available. Ultrafast is claimed for the coming days, up to 8× faster token generation versus standard speed in Codex. Vendor eval excerpts (labeled vendor claim, not re-run here): DeepSWE v1.1 matches Astra at about one-fifth the cost and eclipses GPT-6 Sol’s best by 6.4 percentage points at lower effort/cost; GDP.pdf is higher than Opus 5.5 with fallbacks at less than half the cost; AutomationBench is 2.2 percentage points above Opus 5.5 at medium effort at about one-third the cost, and +4.8 percentage points versus GPT-6 Sol (page wording); OSWorld 2.0 offline partial is +7 percentage points versus GPT-6 Sol at max effort at less than half the cost, within 2.1 percentage points of Astra; Terminal-Bench Science 0.1 more than doubles GPT-6 Sol at max effort at less than half the cost per task, while Astra remains highest on the hardest science at 68.1%; hard-prompt factuality at low effort drops the share of responses with a factual error from 11.4% to 7.7% versus GPT-6 Sol. Third-party board Artificial Analysis https://artificialanalysis.ai/models/gpt-6-1-sol (GPT-6.1 Sol (max), evening GET 200): Intelligence Index 52 (#10/222), 67.8 tok/s (#107/222), $2/$10, Cache Discount 95%, $0.72 per task, 1M context, released 2026-09-29. AA article https://artificialanalysis.ai/articles/gpt-6-1-sol-replaces-gpt-6-sol-after-just-7-days-with-near-astra-intelligence says Sol was replaced by 6.1 Sol after 7 days, about 1 point below Astra II at under a quarter of the cost per task. HN item 49896586, about 974pts on Algolia at the evening check, 2026-09-29T17:06:45Z / about 2026-09-30 01:06 CST. System card addendum https://deploymentsafety.openai.com/gpt-6-1-sol returned HTTP 200 in the evening; its evals were not checked line by line. Morning shipped Cases + HTML only and explicitly skipped generic DevDay news for A/B/C, so Models did not yet have gpt-6-1-sol; evening adds it. Featured, 4 stars for curation quality. Cross-linked with gpt-6-astra, astra-openai-showcase, and claude-sonnet-5-5; gpt-6-astra and claude-sonnet-5-5 relatedModelIds updated. Preview is a Playwright capture of the official page above the fold, cropped to 3:2 WebP. Skipped: thin Show HN pieces (about 1–3pts) TabX / Rhubarb / Instapath / Moching / Fab; HTML Space / Enigma / Scissor already shipped this morning; Agent UI had no solid confirmed-free Community newcomer — not padded; Cases had no new Grok Bot guide. models 17→18 (1.1.12→1.1.13). No cases / html-items / agent-ui edits.
早间收录：Cases +1（Grok Bot for Marketing）、HTML +3（Space Now / Enigma.design / Scissor）Morning catalog: Cases +1 (Grok Bot for Marketing), HTML +3 (Space Now / Enigma.design / Scissor)
Cases：收 grok-bot-for-marketing —— Josh Kim 2026-09-24 官方 Guides https://x.ai/bot/guides/grok-bot-for-marketing 。六席专科营销 Bot（Researcher / Product Marketer / Performance Marketer / Website Ops / Analyst / Project Manager），用虚构航司 X-Air 跑通研究→Docs 简报→Google Ads 壳→营销站 PR→记分牌；席位稳定后再 Template，Ads 写与花钱人审。补齐已收录 GTM / PMs / Engineering 同系列缺口。HTML：收 space-bl2（Show HN item 49898778，Algolia 早间核对 97pts，2026-09-29T19:08:01Z）https://space.bl2.net/ —— 实时三维太阳系/地球卫星云，页内约 52.7 万天体 + CelesTrak 卫星目录。收 enigma-design（item 49896757，核对 31pts，2026-09-29T17:15:56Z）https://enigma.design/ —— 可互动三维恩尼格玛教学。收 scissor-studio（item 49875308，核对 134pts，2026-09-28T08:55:31Z）https://scissor.studio/ —— 免费浏览器矢量+像素编辑器（Rust/WASM，作者尚未开源）。三条均配 Playwright 实拍 3:2 自托管预览。Agent UI：扫 ThreeUI browse 与 Show HN three.js 件后，无足够「可复制给 Agent」且确认 Community 免费的新件，未硬凑。跳过：Wave（item 49843746，9/25，窗口偏早）；Museum of Numbers / Dice Roller（内容站或热度过低）；Jeeves / Destroy Stickman / Rename.Tools 等已在库；芯片/监管/讣闻与 DevDay 通用新闻不进 A/B/C；X @bot 公共镜像 403/空时间线，未硬打 API。cases 131→132（1.4.34→1.4.35）；html-items 100→103（1.3.49→1.3.50）。未改 agent-ui。Cases: added grok-bot-for-marketing — Josh Kim’s official 2026-09-24 Guides piece at https://x.ai/bot/guides/grok-bot-for-marketing. Six specialist marketing Bots (Researcher / Product Marketer / Performance Marketer / Website Ops / Analyst / Project Manager) walk a fictional airline X-Air from research through Docs briefs, Google Ads shells, a marketing-site PR, and a scoreboard; template seats only after they stabilize; Ads writes and spend stay human-gated. Fills the gap next to already-listed GTM / PMs / Engineering guides. HTML: added space-bl2 (Show HN item 49898778, 97pts on Algolia at the morning check, 2026-09-29T19:08:01Z) at https://space.bl2.net/ — real-time 3D Solar System / Earth satellite cloud (~527k bodies + CelesTrak sats). Added enigma-design (item 49896757, 31pts, 2026-09-29T17:15:56Z) at https://enigma.design/ — interactive 3D Enigma lesson. Added scissor-studio (item 49875308, 134pts, 2026-09-28T08:55:31Z) at https://scissor.studio/ — free browser vector+pixel editor (Rust/WASM; open source undecided). All three ship with Playwright-captured 3:2 self-hosted previews. Agent UI: after scanning ThreeUI browse and Show HN three.js pieces, no solid free “copyable for agents” Community newcomer — not padded. Skipped: Wave (item 49843746, Sep 25, outside the window); Museum of Numbers / Dice Roller (content site or too thin); Jeeves / Destroy Stickman / Rename.Tools already catalogued; chip/regulation/obituary and generic DevDay news stay out of A/B/C; X @bot public mirrors 403/empty timeline, API not hammered. cases 131→132 (1.4.34→1.4.35); html-items 100→103 (1.3.49→1.3.50). No agent-ui edits.
最新模型：新开 /models/claude-sonnet-5-5/ —— Anthropic 于 2026-09-28 发布的 Claude 5.5 家族第二款。官方页 https://www.anthropic.com/claude-sonnet-5-5（标题 “Introducing Claude Sonnet 5.5”）晚间 GET HTTP 200。官方称相对 Claude Sonnet 5 为明确升级，生成快 30%+，多数工作花费可低至 30%；价目与 Sonnet 5 同为输入 $2 / 输出 $10 / cache reads $0.20 每百万 token，因 token 更少所以更便宜。定位为 Opus 5.5 的更快、更低成本补位。官方表节选（列 Sonnet 5.5 / Sonnet 5 / Opus 5.5 / GPT-6 Sol）：Terminal-Bench 4.0 70.6% / 10.3% / 66.4%¹ / —；CursorBench 4.0 55.5% / 34.1% / 57.8% / —；GDPval-AA v2.1 1844 / 1449 / 1846 / 1487⁴。页内写 Opus 5.5 在复杂、开放式、需持续判断的工作上仍明显更强，故不把 Sonnet 写成全面第一。平台模型 id：claude-sonnet-5-5；称已在全平台可用（含 AWS / Google Cloud / Azure）。第三方榜 Artificial Analysis https://artificialanalysis.ai/models/claude-sonnet-5-5（变体 Adaptive Reasoning, Max Effort, Default Fallback，晚间 GET 200）：Intelligence Index 56，输出 138.7 tok/s（#29/216），价 $2/$10、Cache Discount 90%、每任务 $7.60，1M 上下文。HN item 49881850（标题 Sonnet 5.5），Algolia 晚间核对 825pts，2026-09-28T17:58:11Z / 约 2026-09-29 01:58 CST。晨间目录未收（当日只收 Jeff + HTML），晚间补上。精选，整理质量 4 星。与 claude-opus-5-5、claude-fable-5-1 互链，并回写 claude-opus-5-5 的 relatedModelIds。新开 /models/jeeves/ —— PostHog 开源、带推理的 Jev 风格 9B 决策模型。HN item 49891290，Algolia 晚间核对 33pts，2026-09-29T11:13:54Z / 约 19:13 CST。仓库 https://github.com/PostHog/jeeves MIT，2026-09-29 gh api 21★；README 写 Inspired by Kev；基于 Qwen3.5-9B（LoRA + pointer head），SFT + CISPO，带 block-4 diffusion drafter；同一请求支持 noul / choice / score，API 与 Jev 同形（/v1/systemone）。权重 https://huggingface.co/PostHog/jeeves。公开仓与 HF 晚间 GET 200。作者表（thinking、greedy、2560-token 上限；Kev-9B / Jev 列为 Kev 公布数字）：Test overall 0.889 vs Kev 0.822 / Jev 0.857；JevBench overall（231 public）0.935 vs Jev 0.866；JevBench hard 0.865 vs Jev 0.730。延迟自述：无 thinking 约 0.3 s/请求，有 thinking 中位约 3.3 s（单 H100）；需 CUDA（Hopper 才用 FP8 kernel）。本库未复跑基准、未代下权重。非 TypeSafe 官方。精选，4 星。与 typesafe-jev、ollaya、laya-system-one 互链并回写；relatedCaseIds 指向晨间 jeff 与 onesie。HTML：收 rename-tools（Show HN item 49887510，Algolia 晚间核对 2pts，2026-09-29T02:54:17Z / 约 10:54 CST）https://rename.tools/en —— 浏览器里用规则链批量改本地文件名：查找替换 / 正则 / 序号 / 大小写 / 插入删除 / 自定义 JS，实时预览冲突；页称 File System Access API 直改（Chrome / Edge 等），其他浏览器可导出 Bash / PowerShell；自称不上传、免费、无需注册。仓库 chenz24/rename.tools 于 2026-09-29 gh search：AGPL-3.0、126★。公开页与仓库 HTTP 200。三条均配 Playwright 实拍 3:2 自托管预览。Agent UI：今日无足够「可复制给 Agent」的免费新件，未硬凑。跳过：Jevstiller（Apache-2.0 distill proxy，与 Jeff/Jeeves 波次重叠，延后以保持本次三条）；systemonemodels.tech（注册表目录，相对可试工具偏薄）；YuliusBox / Pic2svg（过薄或托管 API 边界）；SideKernel / Overmux / Lemma / Sliderino（1–5pts，偏小众或未核实免费 Agent UI）；DevDay 发布尚未开场（keynote 10:00 PT 仍在上海晚间之前）；X 搜索 client-not-enrolled。models 15→17（1.1.11→1.1.12）；html-items 99→100（1.3.48→1.3.49）。未改 cases / agent-ui。Models: new card /models/claude-sonnet-5-5/ — Anthropic’s second model in the Claude 5.5 family, announced 2026-09-28. Official page https://www.anthropic.com/claude-sonnet-5-5 (title “Introducing Claude Sonnet 5.5”) returned HTTP 200 on an evening GET. Vendor claim: a clear upgrade over Claude Sonnet 5, 30%+ faster generation and up to 30% less cost for most work, at the same list price as Sonnet 5 — $2 input / $10 output / $0.20 cache reads per 1M tokens — because it typically uses fewer tokens. Positioned as a faster, lower-cost complement to Opus 5.5. Vendor table excerpt (columns Sonnet 5.5 / Sonnet 5 / Opus 5.5 / GPT-6 Sol): Terminal-Bench 4.0 70.6% / 10.3% / 66.4%¹ / —; CursorBench 4.0 55.5% / 34.1% / 57.8% / —; GDPval-AA v2.1 1844 / 1449 / 1846 / 1487⁴. The page says Opus 5.5 remains clearly stronger on complex, open-ended work that needs sustained judgment, so Sonnet is not framed as first everywhere. Platform model id claude-sonnet-5-5; claimed available on all platforms including AWS / Google Cloud / Azure. Third-party board Artificial Analysis https://artificialanalysis.ai/models/claude-sonnet-5-5 (variant Adaptive Reasoning, Max Effort, Default Fallback, evening GET 200): Intelligence Index 56, 138.7 tok/s (#29/216), $2/$10, Cache Discount 90%, $7.60 per task, 1M context. HN item 49881850 (title Sonnet 5.5), 825pts on Algolia at the evening check, 2026-09-28T17:58:11Z / about 2026-09-29 01:58 CST. Morning catalog skipped it (Jeff + HTML only); this evening wave picks it up. Featured, 4 stars for curation quality. Cross-linked with claude-opus-5-5 and claude-fable-5-1; claude-opus-5-5 relatedModelIds updated. New card /models/jeeves/ — PostHog’s open reasoning Jev-style 9B decision model. HN item 49891290, 33pts on Algolia at the evening check, 2026-09-29T11:13:54Z / about 19:13 CST. Repo https://github.com/PostHog/jeeves MIT, 21★ on the 2026-09-29 gh api; README says Inspired by Kev; Qwen3.5-9B (LoRA + pointer head), SFT + CISPO, plus a block-4 diffusion drafter; the same request supports noul / choice / score via a Jev-shaped /v1/systemone API. Weights at https://huggingface.co/PostHog/jeeves. Public repo and HF evening GET 200. Author table (thinking, greedy, 2560-token cap; Kev-9B / Jev columns are Kev-published): Test overall 0.889 vs Kev 0.822 / Jev 0.857; JevBench overall (231 public) 0.935 vs Jev 0.866; JevBench hard 0.865 vs Jev 0.730. Latency claim: about 0.3 s/request without thinking, ~3.3 s median with it on one H100; CUDA required (Hopper for the FP8 kernel). Benchmarks not re-run here; weights not downloaded. Not TypeSafe official. Featured, 4 stars. Cross-linked with typesafe-jev, ollaya, and laya-system-one, and those cards updated; relatedCaseIds point at morning jeff and onesie. HTML: added rename-tools (Show HN item 49887510, 2pts on Algolia at the evening check, 2026-09-29T02:54:17Z / about 10:54 CST) at https://rename.tools/en — bulk-rename local files in the browser with rule chains (find/replace, regex, sequences, case, insert/remove, custom JS) and a live conflict preview; the page says Chromium browsers rename via the File System Access API and others can export Bash / PowerShell; claims no uploads, free, no signup. Repo chenz24/rename.tools on the 2026-09-29 gh search: AGPL-3.0, 126★. Public page and repo HTTP 200. All three ship with Playwright-captured 3:2 self-hosted previews. Agent UI: no solid free “copyable for agents” new piece today — not padded. Skipped: Jevstiller (Apache-2.0 distill proxy, overlaps the Jeff/Jeeves wave, deferred to keep this set at three); systemonemodels.tech (a registry directory, thin next to a tryable tool); YuliusBox / Pic2svg (thin or a hosted-API boundary); SideKernel / Overmux / Lemma / Sliderino (1–5pts, niche or an unverified free Agent UI fit); DevDay launches not yet (keynote 10:00 PT still ahead of Shanghai evening); X search client-not-enrolled. models 15→17 (1.1.11→1.1.12); html-items 99→100 (1.3.48→1.3.49). No cases / agent-ui edits.
最新模型：新开 /models/ember-1/ —— Fireworks Research 专用模型 Ember-1。官博 https://fireworks.ai/blog/ember-1（PUBLISHED 9/23/2026）与模型页 https://fireworks.ai/models/fireworks/ember-1 晚间 GET 均为 HTTP 200；model path accounts/fireworks/models/ember-1。官方自述：基于 Kimi K3，约少 40% token 且质量相当；Serverless $3 / $0.30 / $15 每百万 token（input / cached input / output）；MoE、2.78T、上下文 1040k；函数调用与图像输入 Supported，模型页 Fine-tuning Not supported（官博另写 training support，能力开关以模型页为准）。HN 首页 item 49868830，Algolia 写入时核对 494pts（研究笔记快照 493pts），2026-09-27T17:31:53Z。Artificial Analysis ember-1 URL HTTP 404，不编第三方榜链接。晨间未收，晚间补上。不打星。Cases：收 onesie（Show HN item 49876472，核对 2pts，2026-09-28T11:37:25Z）https://github.com/frodi-karlsson/onesie — MIT Unix 风格 CLI，给 System One 模型（TypeSafe Jev 直连或 OpenRouter，或 Berget）做管道、退出码门禁、calibrate 与 ONESIE_MOCK；2026-09-28 gh api 0★，不虚构星数。详情含 landing。requiresApproval：不要把密钥或生产命令送进模型门禁。HTML：收 gpu-font（item 49875638，核对 2pts，2026-09-28T09:49:48Z）https://dy.github.io/gpu-font/ —— 浏览器里从裁切的一行字认字体，WebGPU 或 CPU，作者称图片不上传；仓库 dy/gpu-font MIT，gh api 3★。README 实验口径前五 92% / 第一 73%（渲染过的 Google Fonts 裁图）标作者自称，非本站基准。三条均配 3:2 自托管预览。跳过：X API 不可用（client-not-enrolled）；RoadRash 的 Jev 说法页面/脚本对不上，未核实；Nodd 页面近乎空白；Agent Office 无 SPDX license；Redthread 渗透 PoC 超出收录范围。models 14→15（1.1.10→1.1.11）；cases 129→130（1.4.32→1.4.33）；html-items 95→96（1.3.46→1.3.47）。未改 agent-ui。Models: new card /models/ember-1/ — Ember-1, a specialized model from Fireworks Research. Blog https://fireworks.ai/blog/ember-1 (PUBLISHED 9/23/2026) and model page https://fireworks.ai/models/fireworks/ember-1 both returned HTTP 200 on an evening GET; model path accounts/fireworks/models/ember-1. Vendor claims: built on Kimi K3, about 40% fewer tokens at comparable quality; serverless $3 / $0.30 / $15 per 1M tokens (input / cached input / output); MoE, 2.78T parameters, 1040k context; function calling and image input Supported, fine-tuning Not supported on the model page (the blog also mentions training support; capability flags follow the model page). HN front-page item 49868830, 494pts on Algolia at write time (research-note snapshot 493pts), 2026-09-27T17:31:53Z. The Artificial Analysis ember-1 URL returned HTTP 404 — no third-party board link invented. Morning skipped it; evening adds it. Unrated. Cases: added onesie (Show HN item 49876472, 2pts at check time, 2026-09-28T11:37:25Z) at https://github.com/frodi-karlsson/onesie — MIT Unix-style CLI for System One models (TypeSafe Jev direct or via OpenRouter, or Berget) with pipes, exit-code gates, calibrate, and ONESIE_MOCK; 0★ on the 2026-09-28 gh api, stars not invented. Detail page includes a landing. requiresApproval: do not send secrets or production commands into the model gate. HTML: added gpu-font (item 49875638, 2pts at check time, 2026-09-28T09:49:48Z) at https://dy.github.io/gpu-font/ — in-browser font match from a cropped text line, WebGPU or CPU, author says the image is never uploaded; repo dy/gpu-font MIT, 3★ via gh api. README experiment figures top-five 92% / first 73% on rendered Google Fonts crops are the author’s claim, not this site’s benchmark. All three ship with 3:2 self-hosted previews. Skipped: X API unavailable (client-not-enrolled); RoadRash’s Jev claim did not match the page/JS and stays unverified; Nodd’s page was nearly empty; Agent Office has no SPDX license; Redthread is an offensive pentest PoC and out of scope. models 14→15 (1.1.10→1.1.11); cases 129→130 (1.4.32→1.4.33); html-items 95→96 (1.3.46→1.3.47). agent-ui unchanged.
早间收录：Cases +1（Leftovers）、HTML +3（Lofi Cities / Shotcandy / Tiny AI Arena）Morning catalog: Cases +1 (Leftovers), HTML +3 (Lofi Cities / Shotcandy / Tiny AI Arena)
Cases：收 leftovers（Show HN item 49861159，核对 2pts）https://github.com/arpwal/leftovers — MIT macOS 菜单栏，按 footprint 清理 coding agent 留下的进程 / MCP / worktree / 可重建缓存；文档站 HTTP 200，gh api 2★。详情含 landing。HTML：收 lofi-cities（item 49869574，核对 154pts）https://loficities.com/ — 像素夜城 + 浏览器现场合成 lofi，免注册；公开页未见 SPDX 自链仓库。收 shotcandy（item 49871097，核对 2pts）https://shotcandy.vercel.app — 浏览器本地美化截图，仓库 btahir/shotcandy MIT、10★。收 tiny-ai-arena（item 49867775，核对 96pts）https://tinyaiarena.com/ — 四模型 8×8 观战；仓库 hp6/ai-arena 无 SPDX、12★，HN 有移动端难用反馈。四条均配 3:2 自托管预览。跳过：X @bot API 不可用（client-not-enrolled）；theCity（Slaymish/theCity，MIT，Agent UI 候选仍仅 3★）；Cartopolis / InkVec（演示入口或托管边界一般）；无新的免费 ThreeUI Community 条目；无新的官方 Grok Bot 指南。cases 128→129（1.4.31→1.4.32）；html-items 92→95（1.3.43→1.3.46）。未改 agent-ui。Cases: added leftovers (Show HN item 49861159, 2pts at check time) at https://github.com/arpwal/leftovers — MIT macOS menu bar that cleans coding-agent leftover processes / MCP / worktrees / rebuildable caches by footprint; docs site HTTP 200, 2★ via gh api. Detail page includes a landing. HTML: added lofi-cities (item 49869574, 154pts) at https://loficities.com/ — pixel night cities with lofi composed live in the browser, no sign-up; no SPDX repo self-linked from the public page. Added shotcandy (item 49871097, 2pts) at https://shotcandy.vercel.app — local-in-browser screenshot beautifier; repo btahir/shotcandy MIT, 10★. Added tiny-ai-arena (item 49867775, 96pts) at https://tinyaiarena.com/ — four-model 8×8 spectating; repo hp6/ai-arena has no SPDX, 12★, with HN notes on hard mobile scrolling. All four ship with 3:2 self-hosted previews. Skipped: X @bot API unavailable (client-not-enrolled); theCity (Slaymish/theCity, MIT, Agent UI candidate still only 3★); Cartopolis / InkVec (weaker demo entry or hosting boundary); no new free ThreeUI Community listing; no new official Grok Bot guide. cases 128→129 (1.4.31→1.4.32); html-items 92→95 (1.3.43→1.3.46). agent-ui unchanged.
最新模型：新开 /models/maniac-mansion-3d/ —— 社区 demo，作者 s-macke 在 Show HN（item 49864738，Algolia 核对 2pts）自述用 GPT-6 Astra + Blender 把 DOS EGA《Maniac Mansion》屏幕重建成可走 3D；live https://s-macke.github.io/maniac-mansion-3d/ 与仓库均 HTTP 200。README 可核：34 个房间空间、Three.js、烘焙光照、门口/梯子交互、非可玩重制（无谜题/背包）。仓库无 SPDX license，gh api 1★；原作美术属 Lucasfilm / Disney，本卡只链公开 walkthrough，不分发 ROM/原画，也不为该 IP 背书。精选，与 gpt-6-astra、astra-openai-showcase 互链（卡上 relatedModelIds 另含 astra-mini-lol）。Cases：收 shipwithmuse（item 49863641，核对 4pts）https://shipwithmuse.live/ —— 独立策展的 Muse 社区作品目录（Errands / Coding / Local models / Games 等），页内自称 1079 builds；官网 HTTP 200。非 Meta Muse 官方站，赞助位不转推，未逐条核 build；详情含 landing（它是什么 / 对照 / 怎么试 / 边界），避免从 X 点进来只看到薄目录卡。HTML：收 nudl（item 49865832，核对 1pt）https://coreho.github.io/nudl/ —— 浏览器里粘贴链接或整段消息，本地去掉 UTM / 跳转包装等 tracking；仓库 Coreho/nudl MIT，gh api 1★；公开页与仓库 HTTP 200。哪些参数算 tracking 以作者规则为准，本站未审计规则表，CLI/托盘未代装。三条均配 3:2 自托管预览。跳过：X API 不可用（client-not-enrolled）；leftovers（arpwal/leftovers，MIT，下波候选）；theCity（Slaymish/theCity，MIT，Agent UI 候选、星少且当日新建）；jev-windows-agent / peekaboolean / swift-newt（Jev 生态，星低或偏实验）；jauvex（Apache-2.0 但星少、产品边界需再核）；Gigantua（挂在 claude.ai/artifact，托管脆弱）；Trail / kenbaiki / clipyouredit（好玩 HTML，今晚优先 AI 相关）；PeerTalk / LightCloud / 各类 SaaS（边界不清或非本站口径）；html-anything（非今日新发）。models 13→14（1.1.9→1.1.10）；cases 127→128（1.4.30→1.4.31）；html-items 91→92（1.3.42→1.3.43）。未改 agent-ui。Models: new card /models/maniac-mansion-3d/ — a community demo. Author s-macke wrote on Show HN (item 49864738, 2pts on Algolia at check time) that GPT-6 Astra + Blender rebuilt DOS EGA Maniac Mansion screens into walkable 3D; live https://s-macke.github.io/maniac-mansion-3d/ and the repo both HTTP 200. README-checkable: 34 room spaces, Three.js, baked lighting, door/ladder interactions, exploration only (no puzzles/inventory). Repo has no SPDX license, 1★ via gh api; original art is Lucasfilm / Disney — this card only links the public walkthrough, redistributes no ROM/artwork, and does not endorse that IP. Featured, cross-linked with gpt-6-astra and astra-openai-showcase (the card’s relatedModelIds also include astra-mini-lol). Cases: added shipwithmuse (item 49863641, 4pts at check time) at https://shipwithmuse.live/ — an independent curated catalog of public Muse builds (Errands / Coding / Local models / Games, and more); the page claimed 1079 builds; site HTTP 200. Not the official Meta Muse site; sponsor slots are not pushed; individual builds were not verified. The detail page includes a landing (what / contrast / try / boundaries) so an X click does not land on a thin directory card. HTML: added nudl (item 49865832, 1pt at check time) at https://coreho.github.io/nudl/ — paste a link or a whole message in the browser and strip tracking such as UTM / redirect wrappers locally; repo Coreho/nudl MIT, 1★ via gh api; public page and repo HTTP 200. Which params count as tracking follows the author’s rules; the rule table was not audited here, and the CLI/tray were not installed. All three ship with 3:2 self-hosted previews. Skipped: X API unavailable (client-not-enrolled); leftovers (arpwal/leftovers, MIT, next-wave candidate); theCity (Slaymish/theCity, MIT, Agent UI candidate, low stars and created the same day); jev-windows-agent / peekaboolean / swift-newt (Jev ecosystem, low stars or experimental); jauvex (Apache-2.0 but low stars, product boundary still needs a check); Gigantua (hosted on claude.ai/artifact, fragile hosting); Trail / kenbaiki / clipyouredit (fun HTML, AI-related items first tonight); PeerTalk / LightCloud / assorted SaaS (unclear boundary or off this catalog’s scope); html-anything (not a new post today). models 13→14 (1.1.9→1.1.10); cases 127→128 (1.4.30→1.4.31); html-items 91→92 (1.3.42→1.3.43). agent-ui unchanged.
9月26日 · 星期六 · 1 条9/26 · Saturday · 1 items
晚间收录：Models +2（Ollaya、Jev 玩宝可梦）、HTML +1（Print Graph Paper）Evening catalog: Models +2 (Ollaya, Jev Plays Pokémon), HTML +1 (Print Graph Paper)
最新模型：新开 /models/ollaya/ —— 本机跑开源决策模型，像 Ollama 一样 pull / serve Laya 等权重并提供 TypeSafe 兼容的 /v1/systemone（HN item 49848269，Algolia 核对 483pts；仓库 ollaya-dev/ollaya Apache-2.0，gh api 296★）。官网与仓库均 HTTP 200。延迟对比（如 Laya 8–10ms、Jev hosted 236–276ms）标官网自述 / 第三方基准，本库未复跑。新开 /models/jev-plays-pokemon/ —— 社区 live demo，用 TypeSafe Jev 决策模型通关《宝可梦 红》（HN item 49845172，核对 214pts；christianmat/jev-pokemon GPL-2.0，gh api 62★）。作者称 harness 只读 RAM、无脚本通关；README 成本口径约 800–1,300 次/小时、约 $1–1.70 / 24h 标作者称，本库未复跑。不附 ROM。两卡精选，与 typesafe-jev、laya-system-one 互链。HTML：收 Print Graph Paper（item 49854203，核对 3pts）https://printgraphpaper.app/ —— 浏览器里选纸型后下载 PDF 或打印，页面自称 57 种、1:1、免费无水印；无开源仓库可核。三条均配 Playwright 实拍 3:2 自托管预览。跳过：X API 不可用（client-not-enrolled）；Ollaya Web UI（item 49855683，无 license、1★，不单独开卡）；UsageBuddy / Botbin / OOMU 过薄或 SaaS 边界不清；Catescape / CoordMap / unspin 以后再说。models 11→13（1.1.8→1.1.9）；html-items 90→91（1.3.41→1.3.42）。未改 cases / agent-ui。Models: new card /models/ollaya/ — run open decision models locally, Ollama-style pull / serve of weights such as Laya with a TypeSafe-compatible /v1/systemone (HN item 49848269, 483pts on Algolia at check time; repo ollaya-dev/ollaya Apache-2.0, 296★ via gh api). Site and repo both HTTP 200. Latency comparisons (e.g. Laya 8–10ms, Jev hosted 236–276ms) are site or third-party claims, not re-run here. New card /models/jev-plays-pokemon/ — a community live demo of TypeSafe’s Jev decision model playing through Pokémon Red (HN item 49845172, 214pts at check time; christianmat/jev-pokemon GPL-2.0, 62★ via gh api). The author says the harness only reads RAM and there is no scripted win path; the README cost claim of about 800–1,300 calls/hour and about $1–1.70 / 24h stays an author figure, not re-run here. No ROM shipped. Both cards are featured and cross-linked with typesafe-jev and laya-system-one. HTML: added Print Graph Paper (item 49854203, 3pts at check time) at https://printgraphpaper.app/ — pick a paper type in the browser, then download a PDF or print; the page claims 57 types, true 1:1 scale, free, and no watermark; no open-source repo to verify. All three ship with Playwright-captured 3:2 self-hosted previews. Skipped: X API unavailable (client-not-enrolled); Ollaya Web UI (item 49855683, no license, 1★ — not a separate card); UsageBuddy / Botbin / OOMU thin or SaaS-unclear; Catescape / CoordMap / unspin deferred. models 11→13 (1.1.8→1.1.9); html-items 90→91 (1.3.41→1.3.42). No cases / agent-ui edits.
9月25日 · 星期五 · 2 条9/25 · Friday · 2 items
晚间收录：Models +1、HTML +2Evening catalog: Models +1, HTML +2
最新模型：新开 /models/claude-opus-5-5/ —— Anthropic 2026-09-22 发布的 Claude Opus 5.5（官方页 HTTP 200）。官方称典型负载比 Opus 5 少 40% 花费；价目为输入 $4 / 输出 $20 / cache reads $0.20（每百万 token，官方称）。第三方榜 Artificial Analysis Intelligence Index 58（#1/211），速度栏 N/A，$4.00/$20.00、缓存折扣 95%、$5.98/任务（AA 榜单，非本库复跑）。HN item 49803892，Algolia 核对 1793pts。与 claude-fable-5-1、step-5-preview 互链；不新开 Cases。HTML：收 The Last Day（item 49841912，核对 16pts；echohive 互动 2.5D 恐龙绘画，页面自称 Opus 5.5 制作）与 TiddlyInstall（item 49841265，核对 3pts；robertsdotpm/tiddlyinstall MIT，gh api 1★），均配 3:2 自托管预览。跳过：Herobrine（0★ / 无 license）、ShowKit（0★）、Lift-and-Coast（过薄）、interactive-html（agent skill，不是免费 HTML 演示）、Radix / Whiteboard（晨间已跳过）、jevper（以后再说）、LaunchVideo 作为 Models 卡（以后再说；今晚以 Opus 官页 + 恐龙 HTML 为准）。X API 不可用（client-not-enrolled）；未再锤接口。models 10→11（1.1.7→1.1.8）；html-items 88→90（1.3.40→1.3.41）。未改 cases / agent-ui。Models: new card /models/claude-opus-5-5/ — Claude Opus 5.5, announced by Anthropic on 2026-09-22 (official page HTTP 200). Vendor claim: 40% less to run than Opus 5 on typical workloads; price table input $4 / output $20 / cache reads $0.20 per million tokens (vendor figures). Third-party board: Artificial Analysis Intelligence Index 58 (#1/211), speed cell N/A, $4.00/$20.00, 95% cache discount, $5.98/task (AA’s figures, not re-run here). HN item 49803892, 1793pts on Algolia at check time. Cross-linked with claude-fable-5-1 and step-5-preview; no new Cases entry. HTML: The Last Day (item 49841912, 16pts; echohive interactive 2.5D dinosaur painting, page self-claims Opus 5.5) and TiddlyInstall (item 49841265, 3pts; robertsdotpm/tiddlyinstall MIT, 1★ via gh api), each with a 3:2 self-hosted preview. Skipped: Herobrine (0★ / no license), ShowKit (0★), Lift-and-Coast (too thin), interactive-html (agent skill, not a free HTML demo), Radix / Whiteboard (already skipped this morning), jevper (later), LaunchVideo as a Models card (later; tonight is the Opus page plus the dinosaurs HTML entry). X API unavailable (client-not-enrolled); did not hammer it. models 10→11 (1.1.7→1.1.8); html-items 88→90 (1.3.40→1.3.41). No cases / agent-ui edits.
晨更：Cases +1、HTML +3Morning catalog: Cases +1, HTML +3
Grok Bot Cases：收 Personal Jarvis（Show HN item 49833292）——Apache-2.0 本机语音编排桌面，agents / routines / MCP；HN 标题写成 Grok Bot 开源替代，本站按相邻栈收录并加厚 landing。HTML：收 Bastardica（item 49823738，核对 454pts；浏览器 Pyodide 混字体）、koi.rest（item 49837006，核对 115pts；多人锦鲤池）、PipeFitter（item 49836677，核对 3pts；MIT 参数化管件，mankyd/pipefitter），均配 3:2 自托管预览。跳过：Whiteboard IDE（偏设计 IDE，非 Agent UI 组件库）、Radix（需 key / 将上 Pro）、Rig / BoxHaven（过新或与 Errand 重叠且星数过低）、AgentRun（通用 agent DSL，非 Grok Bot 实操）、ThreeUI Structure Flow 变体（父项已在 Agent UI）。X API 不可用（client-not-enrolled）；@bot 未再锤接口。cases 126→127（1.4.29→1.4.30）；html-items 85→88（1.3.39→1.3.40）。未改 agent-ui。Grok Bot Cases: added Personal Jarvis (Show HN item 49833292) — Apache-2.0 local voice-orchestrated desktop with agents / routines / MCP; HN title frames a Grok Bot alternative, catalogued here as an adjacent stack with a thickened landing. HTML: Bastardica (item 49823738, 454pts; in-browser Pyodide font mixer), koi.rest (item 49837006, 115pts; shared koi pond), PipeFitter (item 49836677, 3pts; MIT parametric pipes, mankyd/pipefitter), each with a 3:2 self-hosted preview. Skipped: Whiteboard IDE (design IDE, not an Agent UI component library), Radix (requires a key / Pro coming), Rig / BoxHaven (too new or overlaps Errand with low stars), AgentRun (generic agent DSL, not a Grok Bot workflow), ThreeUI Structure Flow variants (parent already in Agent UI). X API unavailable (client-not-enrolled); did not hammer @bot. cases 126→127 (1.4.29→1.4.30); html-items 85→88 (1.3.39→1.3.40). agent-ui unchanged.
9月24日 · 星期四 · 2 条9/24 · Thursday · 2 items
晚间收录：Cases +1、HTML +2Evening catalog: Cases +1, HTML +2
Grok Bot Cases：收何居上（@hejushang8）2026-09-23 短帖及同日长文——美区 PayPal 绑国内 Visa（截图尾号 2852）再挂美区 Apple ID；口径同 Saily 工具笔记，详情加厚（它是什么 / 对照 / 怎么试 / 边界），requiresApproval；作者自述，本站未开户未绑卡。HTML：收 Show HN CMS-EP-SFX 气隙自解密 HTML（item 49827375，核对 6pts；ApelegHQ/ts-cms-ep-sfx Apache-2.0，gh api 20★）与 C3 Playground（item 49828605，核对 5pts；c3lang/c3c MIT，gh api 5834★），均配 3:2 自托管预览。跳过：XHTML 卡牌站（玩法偏收集、非直接 HTML 片段）、3d-data-globe（license API 空）、Mercury 2.5（缺厂商官页未进 Models）、Jev CAPTCHA 文（typesafe-jev 已在 Models）。X API 不可用，主源用 fxtwitter 镜像核对 PayPal 帖。cases 125→126（1.4.28→1.4.29）；html-items 83→85（1.3.38→1.3.39）。未改 agent-ui / models。Grok Bot Cases: added He Jushang (@hejushang8)’s 23 Sept 2026 short post and same-day article — US PayPal linked to a China-issued Visa (screenshot ending 2852) then to a US Apple ID; same framing as the Saily tool note, thickened landing (what / contrast / try / boundaries), requiresApproval; author self-report, this site did not open an account or bind a card. HTML: added Show HN CMS-EP-SFX air-gapped self-decrypting HTML (item 49827375, 6pts; ApelegHQ/ts-cms-ep-sfx Apache-2.0, 20★ via gh api) and C3 Playground (item 49828605, 5pts; c3lang/c3c MIT, 5834★ via gh api), both with 3:2 self-hosted previews. Skipped: XHTML card game (collectible loop, not pasteable HTML bits), 3d-data-globe (empty license API), Mercury 2.5 (no vendor page → not Models), Jev CAPTCHA write-up (typesafe-jev already on Models). X API unavailable; PayPal post checked via fxtwitter mirror. cases 125→126 (1.4.28→1.4.29); html-items 83→85 (1.3.38→1.3.39). No agent-ui / models edits.
工作日晨更：HTML +3Weekday morning: HTML +3
HTML：收 Show HN 3d-retro（item 49824144，核对 3pts）——CC0 单文件可粘贴复古 3D/Canvas 实验博物馆（站点自称 31，含 /llms.txt 与 API），已配 3:2 自托管预览；收 Fractalysis（item 49816840，核对 2pts）——浏览器 Mandelbrot 无限放大查看器，已配 3:2 预览；收 Simten NAND（item 49817114，核对 3pts）——Apache-2.0 TypeScript HDL 浏览器关卡（simtenHQ/simten，gh api 18★），已配 3:2 预览。本次未改 Grok Bot / Agent UI（无已核实的新免费 ThreeUI；uilib.dev 仓库无 license API 结果未收；Cases 无新可复现 Grok Bot 用法教程）。X @bot 未走 API。html-items 80→83（1.3.37→1.3.38）。HTML: added Show HN 3d-retro (item 49824144, 3pts at check time) — a CC0 single-file pasteable retro 3D/Canvas experiment museum (site claims 31, with /llms.txt and an API) with a 3:2 self-hosted preview; Fractalysis (item 49816840, 2pts) — a browser Mandelbrot infinite-zoom viewer with a 3:2 preview; and Simten NAND (item 49817114, 3pts) — an Apache-2.0 TypeScript HDL browser level (simtenHQ/simten, 18★ on the gh api) with a 3:2 preview. No Grok Bot / Agent UI edits (no verified new free ThreeUI; uilib.dev had no license via the gh api and was skipped; no new reproducible Grok Bot usage tutorial for Cases). X @bot API skipped. html-items 80→83 (1.3.37→1.3.38).
9月23日 · 星期三 · 3 条9/23 · Wednesday · 3 items
晚间收录：Agent UI +1、HTML +2Evening catalog: Agent UI +1, HTML +2
Agent UI：收 Show HN RondoFlow（item 49812163，Algolia 核对 3pts）——MIT 本地优先，React Flow 画布编排本机 Claude Code CLI agent（Director / Planner / Advisor；可混 Claude Code / OpenAI / Perplexity），gh api 18★，无 npm 包，已配 3:2 自托管预览。HTML：收 lineselect（item 49813195，Algolia 核对 2pts）——MIT Rust CLI，管道里 Space 勾选 stdin 行再 Enter 继续，GH Pages 文档页已配 3:2 预览；收 OnDevice Tools（item 49813705，Algolia 核对 1pt）——站点与 README 自称 64 个纯浏览器工具、gh api 为 AGPL-3.0（仓库描述仍写 63），页面含广告与 gtag，处理按作者称在本地，已配 3:2 预览。今晚未改 cases / models。X API 不可用，未走 X；扫 HN Show 与 Algolia（晨间条目之后）。agent-ui 56→57（1.4.20→1.4.21），html-items 78→80（1.3.36→1.3.37）。Agent UI: added Show HN RondoFlow (item 49812163, 3pts on Algolia at check time) — MIT, local-first React Flow canvas that orchestrates local Claude Code CLI agents (Director / Planner / Advisor; can mix Claude Code / OpenAI / Perplexity), 18★ on the gh api, no npm package, with a 3:2 self-hosted preview. HTML: added lineselect (item 49813195, 2pts on Algolia at check time) — MIT Rust CLI that Space-selects stdin lines in a pipeline and Enter-continues, GH Pages docs with a 3:2 preview; and OnDevice Tools (item 49813705, 1pt on Algolia at check time) — site and README claim 64 browser-only tools, gh api license AGPL-3.0 (repo description still says 63), page includes ads and gtag, processing claimed client-side, with a 3:2 preview. No cases / models edits tonight. X API unavailable — X skipped; scanned HN Show and Algolia (after the morning items). agent-ui 56→57 (1.4.20→1.4.21), html-items 78→80 (1.3.36→1.3.37).
午间：Errand 落地页加厚Midday: thicker Errand landing
Grok Bot：加厚 Errand（errand-opensource）详情，避免从 X 点进来只看到目录卡复读。页面改为「它是什么 / 对照 / 怎么试 / 边界」；上手路径改成三条外链（runerrand.dev、runta-dev/errand、HN 49803044 与作者长文），不再用通用 Skill / Routine 模板。预览为作者长文界面图裁成 3:2 自托管。事实对照 HN、仓库 Apache-2.0、官网 FAQ 与 2026-09-22 长文；约 4,000 行 TypeScript 仍标作者称，未新造指标。README 补了「可从 X 链出的案例」最低线。cases 仍为 125，数据版本 1.4.27→1.4.28。未改其他库。Grok Bot: thickened the Errand (errand-opensource) detail page so an X click does not land on a directory card that repeats itself. The page is now what-it-is / contrast / try-it / boundaries; the start path is three outbound links (runerrand.dev, runta-dev/errand, HN 49803044 and the author write-up) instead of the generic Skill / Routine template. Preview is the author write-up UI, cropped to 3:2 and self-hosted. Facts checked against HN, the Apache-2.0 repo, the site FAQ, and the 2026-09-22 write-up; the ~4,000 lines of TypeScript stay labeled as the author’s claim, with no new metrics invented. README now states the minimum bar for cases that may be linked from X. cases stay at 125, dataset 1.4.27→1.4.28. No other libraries edited.
工作日晨更：Grok Bot +1、HTML +2Weekday morning: Grok Bot +1, HTML +2
Grok Bot：收 Errand（HN item 49803044）——Apache-2.0 开源「每席一电脑」桌面客户端（runta-dev/errand / runerrand.dev），作者长文说明依赖 Runta runtime。HTML：收 gpu-lexer（HN item 49619464）——约 27KB 语言无关 WebGPU 语法高亮演示，已配 3:2 自托管预览；收 runboli.com Halo 3 Guardian 个人站（HN item 49732259），已配 3:2 自托管预览。本次未改 agent-ui。X @bot 未走 API；扫 HN 首页/Show 与 Algolia（Grok Bot / CSS demo / three.js / single-file HTML / WebGPU）。cases 124→125，html-items 76→78。Grok Bot: added Errand (HN item 49803044) — Apache-2.0 open-source per-seat-computer desktop client (runta-dev/errand / runerrand.dev); author write-up explains the Runta runtime dependency. HTML: added gpu-lexer (HN item 49619464) — ~27KB language-agnostic WebGPU syntax-highlighter demo with a 3:2 self-hosted preview; and runboli.com Halo 3 Guardian personal site (HN item 49732259) with a 3:2 self-hosted preview. No agent-ui edits this morning. X @bot API skipped; scanned HN front page / Show and Algolia (Grok Bot / CSS demo / three.js / single-file HTML / WebGPU). cases 124→125, html-items 76→78.
9月22日 · 星期二 · 2 条9/22 · Tuesday · 2 items
晚间收录：Models +1、HTML +2Evening: Models +1, HTML +2
最新模型：新开 /models/mimo-v2-6/ —— 小米 MiMo-V2.6 系列旗舰 MiMo-V2.6-Pro-RL（官方页 meta「frontier intelligence, all the modalities, built in public」）；HF 卡（MIT）官方称稀疏 MoE 1.02T / 42B 激活、原生全模态、1M 上下文；同系列 Flash-RL / Distill-Qwen-9B。第三方榜 Artificial Analysis Intelligence Index 46（AA 榜单，非本库复跑）。官方源 mimo.xiaomi.com/mimo-v2-6 + HF，产品源 API 平台，社区源 AA 与 HN（49792730 / 49796660）。HTML：收 Show HN WebGCM（item 49793545）——浏览器 WebGPU 全球气候模型（教育玩具框架，不声称科学精度），已配 3:2 自托管预览；收 System Design Atlas（item 49799241）——带交互架构图的系统设计阅读站（GitHub MIT），已配 3:2 自托管预览。models 9→10（1.1.7），html-items 74→76（1.3.35）。今晚未改 cases / agent-ui。X API 不可用，未走 X；扫 HN 首页与 Show。Models: new card /models/mimo-v2-6/ — Xiaomi MiMo-V2.6 series flagship MiMo-V2.6-Pro-RL (official-page meta “frontier intelligence, all the modalities, built in public”); HF card (MIT) vendor claims sparse MoE 1.02T / 42B activated, native omnimodal, 1M context; series siblings Flash-RL / Distill-Qwen-9B. Third-party board Artificial Analysis Intelligence Index 46 (AA’s figure, not re-run here). Official sources: mimo.xiaomi.com/mimo-v2-6 + HF; product: API platform; community: AA and HN (49792730 / 49796660). HTML: added Show HN WebGCM (item 49793545) — browser WebGPU global climate model (educational-toy framing; no scientific-accuracy claim) with a 3:2 self-hosted preview; and System Design Atlas (item 49799241) — system-design reading with interactive architecture diagrams (GitHub MIT) with a 3:2 self-hosted preview. models 9→10 (1.1.7), html-items 74→76 (1.3.35). No cases / agent-ui edits tonight. X API unavailable — X skipped; scanned HN front page and Show.
Grok Bot：收 HN「Grok Bot」帖下 jjcm 可复现的多席面料询价出样流程（item 49263241 / 父帖 49261514）；收 Show HN Open Bot（item 49365575）——MIT 开源「每席一电脑」相邻栈。HTML：HN 首页 Transformer Explainer（item 49792342）浏览器内 GPT-2 可视化，已配 3:2 自托管预览。Agent UI：Facet（MIT copy-paste R3F 组件库 + MCP / llms.txt），已配 3:2 自托管预览。X @bot 网页 403，未走 X API；扫 HN 首页/Show 与 Algolia（Grok Bot / CSS demo / three.js / single-file HTML）。cases 122→124（1.4.26），html-items 73→74（1.3.34），agent-ui 55→56（1.4.20）。Grok Bot: added jjcm’s reproducible multi-Bot fabric-sourcing flow from the HN “Grok Bot” thread (item 49263241 / parent 49261514), plus Show HN Open Bot (item 49365575) — an MIT open-source per-seat-computer adjacent stack. HTML: HN front-page Transformer Explainer (item 49792342), in-browser GPT-2 visualization with a 3:2 self-hosted preview. Agent UI: Facet (MIT copy-paste R3F component library + MCP / llms.txt) with a 3:2 self-hosted preview. X @bot web returned 403 — no X API calls; scanned HN front page / Show and Algolia (Grok Bot / CSS demo / three.js / single-file HTML). cases 122→124 (1.4.26), html-items 73→74 (1.3.34), agent-ui 55→56 (1.4.20).
9月21日 · 星期一 · 2 条9/21 · Monday · 2 items
晚间收录：HTML +2Evening: HTML +2
HTML：收 Show HN 当日条目 GeoJSON Viewer（item 49784845）——浏览器内粘贴 GeoJSON，按 RFC 7946 做语法+语义校验（闭合环、右手定则、经纬范围、声明 bbox 是否包住数据），错误不画图、警告仍可画，解析/校验/渲染均本地、不上传；收 Show HN GSB（item 49785777）——浏览器商业 trivia，读当年标题把 4 家公司按单年营收从高到低排成金/银/铜，3 次机会，站点自述免费、周一/三/五出新题、可查 archive。两条均配 Playwright 实拍 3:2 自托管预览。html-items 71→73（1.3.33）。今晚未改 Grok Bot / Agent UI / Models。HTML: added same-day Show HN GeoJSON Viewer (item 49784845) — paste GeoJSON in-browser and validate against RFC 7946 (closed rings, right-hand rule, lon/lat ranges, declared bbox containment); errors clear the map, warnings still draw; parse/validate/render stay local with no upload — and Show HN GSB (item 49785777), a browser business trivia game that ranks four companies by single-year revenue Gold→Silver→Bronze with three tries; the site says it is free, with new puzzles Mon/Wed/Fri and an archive. Both ship with Playwright-captured 3:2 self-hosted previews. html-items 71→73 (1.3.33). No Grok Bot / Agent UI / Models edits tonight.
Grok Bot：收 6yuwei《共享云电脑与六种可测工作流》，以及 Cursor 官方文档《Get started with Grok Bot》。HTML：Show HN Scrolling Text Generator（item 49778087）浏览器跑马灯/字幕导出，已配 3:2 自托管预览。Agent UI：ThreeUI Community Gallery（community-sync 确认免费）。X @bot API 未强打，改走网页/公开源。Grok Bot: added 6yuwei’s shared-cloud-computer and six-workflow guide, plus Cursor’s official Get started with Grok Bot. HTML: Show HN Scrolling Text Generator (item 49778087) with a self-hosted 3:2 preview. Agent UI: ThreeUI Community Gallery (community-sync confirmed free). Skipped hard X @bot API calls; used web/public sources.
最新模型：新开 /models/step-5-preview/ —— 阶跃星辰 StepFun 当日（9/20）发布的新旗舰 Step 5 Preview，官方定位面向 agentic 工作（软件工程、Agent、专业知识工作、金融）；官方称稀疏 MoE 600B 总参 / 27B 活跃、1M 上下文、原生图像输入，当日经产品与 API 开放，开源权重定档 10-15。第三方榜 Artificial Analysis Intelligence Index 44、$1 / $2.70 每百万 token 均标「AA 榜单」，本库未复跑任何基准；Hugging Face 的 stepfun-ai/Step-5-Preview-BF16 今晚返回 401 / 空仓，不列为来源，权重尚不可下载。官方源为 stepfun.com 公告页，社区源为 AA、OrcaRouter 博文（权重仓占位旁证）与 HN（item 49772532），四条 URL 均核到 200；与 GPT-6 Astra、Claude Fable 5.1 两卡按官方对照表互设 relatedModelIds。预览为 Playwright 实拍官方页首屏。HTML：收 HN 当日爆帖 AI 活动海报（item 49764791，约 1600+ 分）—— John Hartnup 2026-06-07 的静态 HTML 长文，逐轮贴提示词与生成图、整理出 Bauhaus / Swiss / Risograph 等可抄风格清单，types 取 docs / other；收 Show HN Sparse-SNN（item 49770006）—— AwareLiquid HF Space 纯前端演示，5% 密度、LIF 神经元、果蝇连接组统计掩码的稀疏脉冲网络在浏览器内识别手写数字（ONNX Runtime Web），页面自述节能倍率与「研究产物」边界照录，不进精选。两条均配 Playwright 实拍 3:2 自托管预览。html-items 68→70（1.3.31），models 8→9（1.1.6）。Agent UI：今晚无已核实的新免费 ThreeUI，未新增；Grok Bot：今晚无新案例（pen.dev 案例走独立 PR）；X API 仍不可用，未走 X。Models: new card /models/step-5-preview/ — Step 5 Preview, StepFun’s new flagship announced today (9/20), pitched by the vendor for agentic work (software engineering, agents, professional knowledge work, finance); vendor claims a sparse MoE with 600B total / 27B active parameters, 1M context, and native image input, available same-day via products and API, with open weights scheduled for 10-15. The Artificial Analysis Intelligence Index of 44 and the $1 / $2.70 per-million-token pricing are labelled as AA board figures; no benchmark was re-run here. The Hugging Face repo stepfun-ai/Step-5-Preview-BF16 returned 401 / empty tonight and is not listed as a source — weights are not downloadable yet. Official source is the stepfun.com announcement page; community sources are AA, the OrcaRouter post (corroborating the placeholder weights repo), and HN (item 49772532); all four URLs returned 200. Cross-linked via relatedModelIds with GPT-6 Astra and Claude Fable 5.1, both of which appear in the vendor’s own comparison table. Preview is a Playwright capture of the official page above the fold. HTML: added today’s HN viral hit AI event posters (item 49764791, ~1,600+ points) — John Hartnup’s 2026-06-07 static HTML essay that shows each prompt round next to its generated poster and distils a copyable style list (Bauhaus / Swiss / Risograph, …), typed docs / other; and Show HN Sparse-SNN (item 49770006) — AwareLiquid’s client-side HF Space demo where a 5%-density LIF spiking network with Drosophila-connectome-statistics masks classifies hand-drawn digits in the browser (ONNX Runtime Web); the page’s energy-saving claims and “research artifact” boundary are recorded as stated; not featured. Both ship with Playwright-captured 3:2 self-hosted previews. html-items 68→70 (1.3.31), models 8→9 (1.1.6). Agent UI: no verified new free ThreeUI tonight, nothing added; Grok Bot: no new case tonight (the pen.dev case rides a separate PR); X API still unavailable — X skipped.
9月19日 · 星期六 · 1 条9/19 · Saturday · 1 items
晚间收录：HTML +2、Models +1Evening catalog: HTML +2, Models +1
HTML：收 Show HN 当日条目 Mineral Gems（item 49759144）——浏览器程序化生成矿物宝石的纯前端互动页（gems.0x0000007a.com）；补收 Show HN Page Rage（item 49746963，2026-09-18）——免费「砸网页」小游戏（page-rage.com）：预制页面模式无需安装、任意浏览器可玩，砸任意网页需装浏览器扩展（扩展不在收录范围）；条目在 9/18 白天、晨扫与昨晚波次均未入库，今晚补齐。两条均配 Playwright 实拍 3:2 自托管预览，types 按 schema 枚举取 interactive / other。最新模型：新开 /models/laya-system-one/ —— ConvAI Innovations 的 Laya，作者称为非自回归、亚 35ms 的开源 System 1 决策模型族（HF 模型页核到 Apache 2.0），提供 choice / score / noul 三类结构化决策与多语路由，对照站内已收录的 TypeSafe Jev（两卡已互设 relatedModelIds）；延迟、基准与 Jev 对比数字一律标「作者称 / 第三方整理」，本库未复跑，作者与 Jev 的「谁先提出」争议不站队。官方源为 HF hub，产品源为作者长文页与 HF Space，社区源为 HN（item 49765348）与 GitHub 仓库，五条 URL 均核到 200；预览为 Playwright 实拍作者页首屏。html-items 66→68（1.3.30），models 7→8（1.1.5）。Agent UI：今晚无已核实的新免费 ThreeUI，未新增；Grok Bot：今晚无新案例；X API 仍不可用，未走 X。HTML: added same-day Show HN Mineral Gems (item 49759144) — a client-side interactive page that procedurally generates mineral gems in the browser (gems.0x0000007a.com) — and caught up on Show HN Page Rage (item 49746963, 2026-09-18), a free “smash the page” browser game (page-rage.com) whose pre-made-pages mode needs no install and runs in any browser, while smashing arbitrary pages requires a browser extension (out of scope here); it landed daytime 9/18 and was missed by both the morning scan and last night’s wave. Both ship with Playwright-captured 3:2 self-hosted previews; types use the schema enums interactive / other. Models: new card /models/laya-system-one/ — Laya from ConvAI Innovations, which the author presents as an open-weight non-autoregressive sub-35ms System 1 decision family (HF model page shows Apache 2.0) with choice / score / noul primitives and multilingual routing, catalogued alongside TypeSafe Jev (both cards now cross-link via relatedModelIds); all latency, benchmark, and Jev-comparison figures are labelled author or third-party claims — not re-run here — and this library takes no side in the author-vs-Jev “who was first” dispute. Official source is the HF hub; product sources are the author’s long-form page and the HF Space; community sources are HN (item 49765348) and the GitHub repo; all five URLs returned 200. Preview is a Playwright capture of the author page above the fold. html-items 66→68 (1.3.30), models 7→8 (1.1.5). Agent UI: no verified new free ThreeUI tonight, nothing added; Grok Bot: no new case tonight; X API still unavailable — X skipped.
9月18日 · 星期五 · 2 条9/18 · Friday · 2 items
晚间收录：HTML +2、Agent UI +1Evening catalog: HTML +2, Agent UI +1
HTML：收 Show HN 当日条目 Green Screen Remover（item 49749795）——浏览器本地抠绿幕/蓝幕视频与图片，不上传、免账号、无水印；补收 Show HN Visual HTML（item 49713864，2026-09-15）——保源码的开源 HTML 可视化编辑器，晨扫遗漏今晚补齐。两条均配 Playwright 实拍 3:2 自托管预览。Agent UI：ThreeUI Community Sylva Living World（community-sync 确认免费；同页 Sakura / Maple / Sequoia 季节变体为 PRO，不收录）。X API 仍 403，未走 X；最新模型未新增（kbai-skill 为 BSL 许可，不入库）；今晚无新 Grok Bot 案例。HTML: added same-day Show HN Green Screen Remover (item 49749795)—in-browser chroma key for green/blue-screen video and images with no upload, account, or watermark—and caught up on Show HN Visual HTML (item 49713864, 2026-09-15), a source-preserving open-source visual HTML editor missed in the morning scan. Both ship with Playwright-captured 3:2 self-hosted previews. Agent UI: ThreeUI Community Sylva Living World (confirmed free in community-sync; the Sakura / Maple / Sequoia seasonal variants on the same page are PRO and not listed). X API still 403—X skipped; no Models addition (kbai-skill is BSL-licensed, not catalogued); no new Grok Bot case tonight.
Grok Bot：收官方 Guides《Designing Grok Bot with Grok Bot》（John Bai：Experiments / Motion God / Figma Bro），以及 bot.store Academy《Teach a task → Skill》。HTML：Show HN Paper Effects（item 49747752）浏览器文字 match cut，已配 3:2 自托管预览。Agent UI：ThreeUI Community Bookshelf（community-sync 确认免费）。X @bot API 403/client-not-enrolled，改走网页/公开源；付费 Substack（Linas）与 Medium member-only 未入库。Grok Bot: added official Guides “Designing Grok Bot with Grok Bot” (John Bai: Experiments / Motion God / Figma Bro) and bot.store Academy “Teach a task → Skill”. HTML: Show HN Paper Effects (item 49747752) in-browser text match cuts, with a 3:2 self-hosted preview. Agent UI: ThreeUI Community Bookshelf (confirmed free in community-sync). X @bot API returned 403/client-not-enrolled—used web/public sources; paid Substack (Linas) and Medium member-only were not catalogued.
Grok Bot：收 DataCamp《Skills / Routines / Approvals》Scout 周计划教程，以及 Developers Digest Routine 反转自动化解读。HTML：Show HN Capsule（item 49712278）单文件 Web 应用，已配 3:2 自托管预览。Agent UI：ThreeUI Community Engraved Certificate（community-sync 确认免费）。X @bot API 403，改走网页/公开源；Galaxy 为活动页未入库。Grok Bot: added DataCamp’s Skills/Routines/Approvals Scout weekly-plan tutorial and Developers Digest’s routines-vs-builders essay. HTML: Show HN Capsule (item 49712278) single-file web apps, with a 3:2 self-hosted preview. Agent UI: ThreeUI Community Engraved Certificate (confirmed free in community-sync). X @bot API returned 403—used web/public sources; Galaxy is an event page and was not catalogued.
最新模型：新开 /models/typesafe-jev/ — TypeSafe AI 9 月 15 日发布的首个公开 System One Model「Jev」，面向软件内决策而非聊天：输入非结构化状态，输出类型安全的结构化决策并附校准置信度；官方称新架构 + 并行采样 + RLCD。目前早期访问（官网 Join Waitlist）。官方源为 TypeSafe 博文与官网，社区源为 HN 帖（item 49717558）。约 70–500ms、$0.042 / MTok 输入、193.6× / 444.6× 等数字一律标官方称，本库未复跑、未实测 API。精选、4 星，无预览图。models 6→7，数据版本 1.1.4。Grok Bot / HTML / Agent UI 晚间无新增。Models: new card /models/typesafe-jev/ — “Jev”, TypeSafe AI’s first public System One Model, announced 15 Sept; built for in-software decisions rather than chat: unstructured state in, type-safe structured decisions out with calibrated confidence; vendor claims a new architecture, parallel sampling, and RLCD. Currently early access (Join Waitlist on the site). Official sources are the TypeSafe blog post and homepage; community source is the HN thread (item 49717558). The ~70–500ms, $0.042 / MTok input, and 193.6× / 444.6× figures are all labelled vendor claims — not re-run here, API untested. Featured, 4 stars, no preview image. models 6→7, dataset 1.1.4. No Grok Bot / HTML / Agent UI additions this evening.
工作日晨更：Grok Bot +2、HTML +1Weekday morning: Grok Bot +2, HTML +1
Grok Bot：增收官方 Guides《Grok Bot for GTM》《Grok Bot for PMs》。HTML：Show HN Redis City（item 49676425）浏览器内 Redis 内部 3D 互动图，已配 3:2 自托管预览。Agent UI：无新增。Grok Bot: added official Guides “Grok Bot for GTM” and “Grok Bot for PMs”. HTML: Show HN Redis City (item 49676425), browser-only interactive 3D Redis internals map, with a 3:2 self-hosted preview. Agent UI: none.
Grok Bot：增收官方 Guides《Grok Bot 101》《Grok Bot for Support》《Templates for Grok Bot》《How I run multiple teams of Grok Bots》，以及社区 Roster Room（82 套小队提示）、Grok Ship（scout/ship 软件工厂；README 注明已被 Firstmate 模板替代）、RongleCat/awesome-grok-bot（中英日目录）。HTML：Show HN Neobrutalism.dev（item 49699159）与 code-preview-element 自渲染代码块。Agent UI：Zephyr（Agent API / Zephyr.agent.act）。均已配 3:2 自托管预览。Grok Bot: added official Guides “Grok Bot 101”, “Grok Bot for Support”, “Templates for Grok Bot”, “How I run multiple teams of Grok Bots”, plus community Roster Room (82 team prompts), Grok Ship (scout/ship factory; README notes Firstmate template supersession), and RongleCat/awesome-grok-bot (EN/ZH/JA directory). HTML: Show HN Neobrutalism.dev (item 49699159) and code-preview-element. Agent UI: Zephyr (Agent API / Zephyr.agent.act). All HTML/Agent UI cards ship 3:2 self-hosted previews.
9月14日 · 星期一 · 3 条9/14 · Monday · 3 items
新开 Image 2.5 提示词看板：首条 img.dsxzai.com 画廊（@dashiAIxz）New Image 2.5 prompts board: first card is the img.dsxzai.com gallery (@dashiAIxz)
新增 /image-prompts/ 看板（数据 data/image-prompts.json，字段见 image-prompts-schema.md，另有 /image-prompts.json），与最新模型看板平行，专收 GPT-Image-2.5 / ChatGPT Images 2.5 的提示词画廊、案例合集与可抄的提示词；已进顶栏、首页左栏、页脚与我的收藏。首条 /image-prompts/dsxzai-image-25-gallery/：@dashiAIxz（大师的AI小灶）9 月 14 日帖称「GPT6 太耀眼了，导致 Image 2.5 虽然能力很强没多少人关注到」，把网上 Image 2.5 案例与提示词整理成 img.dsxzai.com（页面标题「生成提示词案例库」），作者称每个案例都有提示词和来源、复制就能生成同款。站点 2026-09-14 核到 200，按模型分栏、按标签筛选，每条列图、完整提示词、复制提示词与查看来源。作者说法与本库核到的分开写，不写点赞浏览，条数只记接口当日自报值并标口径。预览取作者 X 帖附图裁 3:2。数据版本 1.0.0 / 1 条。Added the /image-prompts/ board (data in data/image-prompts.json, fields in image-prompts-schema.md, plus /image-prompts.json), parallel to the Models board and dedicated to prompt galleries, case roundups, and copyable prompts for GPT-Image-2.5 / ChatGPT Images 2.5; wired into the header, home rail, footer, and My favorites. First card, /image-prompts/dsxzai-image-25-gallery/: on 14 Sept @dashiAIxz (大师的AI小灶) posted that “GPT-6 is so dazzling that Image 2.5, strong as it is, got little attention,” and collected Image 2.5 cases and prompts into img.dsxzai.com (page title 「生成提示词案例库」); the author says every case has its prompt and source and “copy it and you get the same result.” The site returned 200 on 2026-09-14, with model tabs, tag filters, and per-case image, full prompt, copy-prompt, and view-source. Author claims and what this library checked are kept apart; no likes or views recorded; the only count is the API’s self-reported total on the day, with its scope stated. Preview is the author’s X screenshot cropped to 3:2. Dataset 1.0.0 / 1 item.
Grok Bot：增收官方 Guides《Grok Bot for Engineering》（x.ai/bot/guides/grok-bot-for-engineering，Lingxi Li，2026-09-10；异于已收录 lingxi-engineering-org X 长文）与 MoClaw《Cursor Cloud Agent vs Grok Bot》（moclaw.ai，文内称 2026-09-11 有修订）。HTML：Show HN Soundmap（item 49644603，soundmap.darebuild.com）与 Hopla Games（item 49682477，hoplagames.com/en）。Agent UI：ThreeUI Community Cloud Field（/backgrounds/portal-field/cloud-field）与 Dot Border Button（/buttons/rectangle-buttons/dot-border-button），均核到 Copy Prompt、无 Pro template 遮罩。Grok Bot: added the official Guides piece “Grok Bot for Engineering” (x.ai/bot/guides/grok-bot-for-engineering, Lingxi Li, 2026-09-10 — distinct from the already-listed lingxi-engineering-org X article) and MoClaw’s “Cursor Cloud Agent vs Grok Bot” (moclaw.ai; article notes a Sep 11, 2026 update). HTML: Show HN Soundmap (item 49644603, soundmap.darebuild.com) and Hopla Games (item 49682477, hoplagames.com/en). Agent UI: ThreeUI Community Cloud Field (/backgrounds/portal-field/cloud-field) and Dot Border Button (/buttons/rectangle-buttons/dot-border-button), both checked for Copy Prompt with no Pro template overlay.
按冰河（@binghe）9 月 13 日帖「一键高级感」的三站收录，来源帖正文无外链，官网与仓库均于 2026-09-14 核到 200。React Bits（reactbits.dev / DavidHDev/react-bits）：React 动画组件浏览站，「160+」为作者自述，仓库许可证为 MIT + Commons Clause；Animata（animata.design / codse/animata）：MIT、copy-paste 的 React + Tailwind 动画组件，「150+」为作者自述；Vengeance UI（帖内写作 VengenceUI，域名 vengenceui.com，仓库 Ashutoshx7/VengeanceUI，MIT）：偏落地页交互的动画组件，作者的营销描述不转成事实数字。三条均精选、类型 component / motion（Vengeance UI 另加 landing），预览为官网现场截图。均非 Grok Bot 案例。HTML 54→57，数据版本 1.3.22。Three sites from 冰河’s (@binghe) 13 Sept “one-click premium feel” post; the post itself carries no external links, and each official site and repo returned 200 on 2026-09-14. React Bits (reactbits.dev / DavidHDev/react-bits): a browse-and-copy site of animated React components, “160+” is the author’s claim, repo licensed MIT + Commons Clause; Animata (animata.design / codse/animata): MIT, copy-paste React + Tailwind animated components, “150+” is the author’s claim; Vengeance UI (spelled VengenceUI in the post, domain vengenceui.com, repo Ashutoshx7/VengeanceUI, MIT): landing-page-leaning animated components, with the author’s marketing copy not turned into factual figures. All three featured, typed component / motion (Vengeance UI also landing), previews are live screenshots of the official sites. None are Grok Bot cases. HTML 54→57, dataset 1.3.22.
9月13日 · 星期日 · 2 条9/13 · Sunday · 2 items
晚间：HTML +2、Agent UI +2Evening: HTML +2, Agent UI +2
HTML 增收 Session Context（Show HN 49678035，白话讲清一页网页未授权前能测到什么）与 Programming Museum（Show HN 49682008，可漫游的 MIT/Unlicense 源码展厅）。Agent UI 增收 ThreeUI Community：Typography Vortex、Semantic Bloom（均核到免费页、无 Pro template 遮罩）。Grok Bot / Models 晚间无新增。HTML: Session Context (Show HN 49678035, plain-English map of what a page can measure before any permission) + Programming Museum (Show HN 49682008, walkable MIT/Unlicense source-code rooms). Agent UI: ThreeUI Community Typography Vortex and Semantic Bloom (both free Community pages, no Pro template overlay). No Grok Bot / Models additions this evening.
按 Hao 的两条线索收录。最新模型：新开社区 showcase 卡 /models/astra-openai-showcase/ — 来源是宝玉（@dotey）9 月 12 日帖「GPT-6 Astra 社区作品集合……或者可以去官网看看 Showcase」，官方源 OpenAI Developers Showcase 页（2026-09-13 核到 200，分 Games / Apps / Websites / API，带 GPT-6 Astra 模型筛选），社区源为 dotey 帖与其引用的 OpenAIDevs 合集文帖；文内作品清单只作官方社区文自述，不复述未核条数。GPT-6 Astra 卡同步加上 Showcase 官方源与 dotey 社区源，relatedModelIds 与新卡互指。models 5→6，数据版本 1.1.3。Grok Bot 库：增收酱紫表（@pengchujin）9 月 13 日帖的 Saily 美国 +1 号 eSIM 工具笔记 — 包年套餐、优惠码 JIANGZ7166 立减 5 刀、约 6.99 美元 / 年，以及「Claude 收不到验证码、ChatGPT / Telegram / PayPal 基本可以」均为作者自述，优惠码归属作者，本站不担保可用性、不构成购买推荐；标为旁路工具笔记，不是 Grok Bot 用法。cases 104→105，数据版本 1.4.19。两张预览分别取 OpenAIDevs 文章封面与作者截图拼图。Two finds from Hao. Models: a new community-showcase card /models/astra-openai-showcase/ — sourced from 宝玉’s (@dotey) 12 Sept post (“a collection of GPT-6 Astra community builds… or browse the Showcase on the official site”); the official source is the OpenAI Developers Showcase page (200 on 2026-09-13; Games / Apps / Websites / API tabs with a GPT-6 Astra model filter), community sources are the dotey post and the OpenAIDevs roundup article post it quotes; the article’s build list is recorded as the official community article’s own claims, with no unverified counts repeated. The GPT-6 Astra card gains the Showcase official source and the dotey community source, and the two cards cross-link via relatedModelIds. models 5→6, dataset 1.1.3. Grok Bot library: added 酱紫表’s (@pengchujin) 13 Sept Saily US +1 eSIM tool note — the annual plan, promo code JIANGZ7166 for $5 off (about $6.99/yr), and “Claude did not receive codes; ChatGPT / Telegram / PayPal mostly did” are all the author’s claims; the code belongs to the author; no availability guarantee and no purchase recommendation; marked as an adjacent tool note, not a Grok Bot playbook. cases 104→105, dataset 1.4.19. Previews: the OpenAIDevs article cover and a composite of the author’s screenshots.
9月12日 · 星期六 · 1 条9/12 · Saturday · 1 items
晚间：HTML +2、Agent UI +2Evening: HTML +2, Agent UI +2
Grok Bot 增收程意（@ChengYi3629）的 native-subtitle-quote-image：保留视频内嵌字幕、精确取帧并生成 3:4 社交长图的开放 Agent Skill（MIT，2026-09-11 核对约 616 star，最新 v2.1.1）。来源推文称升级 2.0、抖音「平均每个视频赞藏 1 千」为作者自称，条目里只作引用不作核验数字；README 是 Codex 装法，Grok Bot 侧按官方 skills 目录放置，本站未实测。预览图取自仓库 assets 图标。Grok Bot: ChengYi’s (@ChengYi3629) native-subtitle-quote-image — an open Agent Skill that keeps burned-in subtitles, grabs exact frames, and renders 3:4 social cards (MIT, ~616 stars checked 2026-09-11, latest v2.1.1). The source post’s “upgraded to 2.0” and “~1k likes+saves per Douyin video” are the author’s claims and are quoted, not verified, in the entry; the README documents the Codex install path, and the Grok Bot placement follows the official skills directory, untested here. Preview image is the repo’s assets icon.
最新模型：收录 Tripo3D 上的 Astra 3D 提示词库（社区资源卡）Models: Astra 3D prompt gallery on Tripo3D (community resource card)
按 Hao 要求收录 @cnyzgkc（木马人）9 月 9 日中文帖「分享一个 GPT-6 Astra 3D 提示词网站，拥有 200+ 的提示词，而且完全免费」。与 Astra 复刻迷你 LOL 同一做法：不塞进 GPT-6 Astra 的来源，而是在模型看板开一张社区资源卡 /models/astra-tripo-3d-prompts/。community 源指作者 X 视频帖（约 28 秒录屏），product 源指 Tripo3D 的 Astra 3D 提示词页（中文路径，2026-09-11 核到可达），official 源仍是 OpenAI Astra 公告，只作所用模型页。收录边界：「200+」「完全免费」是发帖人自称，录屏封面帧显示 201 条、当日页面显示 215 条，均未逐条复数；Tripo3D 是第三方 3D 生成产品的提示词画廊，页面带其产品入口，本库不背书、未测试账号或付费门槛。页面里也收有岚叔与李岳的「复刻英雄联盟」提示词，并自注为推荐内容、非确认输入，与本库 demo 卡边界一致。预览取发帖人录屏封面（Tripo3D 页首屏）。不打星。GPT-6 Astra 卡的 relatedModelIds 加上本卡，两卡互指。models 4→5，数据版本 1.1.2。At Hao’s request, catalogued @cnyzgkc’s (木马人) 9 Sept Chinese post: “sharing a GPT-6 Astra 3D prompt site with 200+ prompts, completely free.” Same treatment as the Astra mini-LoL demo: instead of stuffing it into GPT-6 Astra’s sources, it gets its own community-resource card on the models board, /models/astra-tripo-3d-prompts/. The community source is the author’s X video post (~28s screen recording); the product source is Tripo3D’s Astra 3D prompts page (Chinese path, confirmed reachable 2026-09-11); the official source stays the OpenAI Astra announcement, as the page for the model used only. Boundaries: “200+” and “completely free” are the poster’s claims; the recording’s poster frame shows 201 and the page showed 215 on the day — none re-counted item by item. Tripo3D is a third-party prompt gallery run by a 3D-generation product, with its own product entry points on the page; not endorsed here, sign-up and paywall gates untested. The page also lists the @LufzzLiz and 李岳 “League of Legends clone” prompts and itself marks them as recommendations, not confirmed inputs — matching this library’s demo-card boundary. Preview is the poster’s recording frame (top of the Tripo3D page). Unrated. GPT-6 Astra’s relatedModelIds gains this card, so the two point at each other. Models 4→5, dataset version 1.1.2.
Grok Bot 增收 Young Urban Project《怎么用 Grok Bot》（真用例 vs 炒作，2026-09-09）与 Every Inc Compound Engineering 工程循环插件（README 写明 Grok Bot 共用 Cursor 插件库）。HTML 增收 DAUB（给 AI 的 CSS 组件库，llms.txt / MCP）与 njX UI（CDN 落地页 CSS）。Agent UI 增收 ThreeUI Community：3D Paper · Original、Woven Cloth · Woven Cloth。X @bot 网页侧未见新的可复现用法帖；HN 首页多为泛新闻，未塞进三库。Grok Bot: Young Urban Project How to Use Grok Bot (real cases vs hype, 2026-09-09) + Every Inc Compound Engineering plugin (README: Grok Bot shares the Cursor plugin library). HTML: DAUB (CSS kit for AI, llms.txt / MCP) + njX UI (CDN landing CSS). Agent UI: ThreeUI Community 3D Paper · Original and Woven Cloth · Woven Cloth. No new reproducible @bot usage posts via the public web pass; HN front page was mostly general news and stayed out of the three libraries.
Models 增收 DeepSeek-V4.1-Flash（官方 2026-09-10 发布，API `deepseek-flash`；基准与架构数字为厂商自述，HN 49639090 只作热度参考）。HTML 增收 Relativity Park（光速 5 km/h 的狭义相对论互动，Show HN 49637385）与 Roadwright（画轮见路，Show HN 49641755）。Agent UI 增收 ThreeUI Community：Matrix Junction · Laser、Liquid Metal Button · Pill。Grok Bot 晚间无新增。Models: DeepSeek-V4.1-Flash (official release 2026-09-10, API `deepseek-flash`; benchmark and architecture figures are vendor claims, HN 49639090 only as an interest pointer). HTML: Relativity Park (special relativity at c = 5 km/h, Show HN 49637385) + Roadwright (draw a wheel, see its road, Show HN 49641755). Agent UI: ThreeUI Community Matrix Junction · Laser and Liquid Metal Button · Pill. No Grok Bot additions this evening.
Grok Bot 增收 awesome-grokbot（661 条可点进 x.ai/bot 分享目录，异于已收录的 Skill/MCP 目录）与 Continuum Routine/Skill 工作指南（第三方上报上限标为未确认）。HTML 增收 Maple 零构建原子 CSS（HN 49612277）、Topological Picture Book 钢笔拓扑曲面（HN 49618050）。Agent UI 增收 ThreeUI Community：PerformanceGauges · Tachometer、CircleButtons · Play。X @bot 网页检索无新的可复现帖。Grok Bot: awesome-grokbot (661 live x.ai/bot shares; distinct from the already-listed skills/MCP directory) + Continuum Routine/Skill working guide (third-party reported ceilings flagged unconfirmed). HTML: Maple zero-build atomic CSS (HN 49612277) + Topological Picture Book pen-and-ink surfaces (HN 49618050). Agent UI: ThreeUI Community PerformanceGauges · Tachometer + CircleButtons · Play. No new reproducible @bot posts via web search.
最新模型：收录 Astra 复刻迷你 LOL（社区 demo 卡）Models: Astra recreates a mini LoL (community demo card)
按 Hao 要求收录 @LufzzLiz（岚叔）9 月 8 日中文帖。没塞进 GPT-6 Astra 的来源里，而是在模型看板开一张社区 demo 卡 /models/astra-mini-lol/：主源指作者 @konstiwohlwend 的 X 视频帖（约 44 秒，「让 Astra 照抄 LoL，它真做出来了」），次源指岚叔中文转述与两步提示词（先做 5 英雄的完整 LOL；再骂回去「廉价山寨」，要先写计划、按真实尺寸机制实现、UI 原生不许 HTML 贴）；official 源仍是 OpenAI Astra 公告，只作所用模型页。作者说提示词在回复里，本库抓不到，故提示词标为岚叔中文转述。收录边界：作者自述 + 短视频，无可玩构建，玩法深度未核实；英雄联盟 IP 属 Riot，只记模型能力展示，不分发游戏文件；不写任何编造数字，不打星。预览取作者视频封面（选人界面 Ahri，裁掉黑边）。同时 models.json 新增可选 relatedModelIds：demo 卡与 GPT-6 Astra 卡互指，详情页多一个「同板相关」区块渲染整卡。models 2→3，数据版本 1.1.0。At Hao’s request, catalogued @LufzzLiz’s 8 Sept Chinese post. Rather than stuffing it into GPT-6 Astra’s sources, it gets its own community-demo card on the models board, /models/astra-mini-lol/: primary source is the author @konstiwohlwend’s X video post (~44s, “asked Astra to copy LoL and it actually did”); secondary is @LufzzLiz’s Chinese write-up with the two-step prompt (first a full LoL-like game with 5 champions; then push back “cheap clone” — plan first, real scale and mechanics, native UI, no HTML overlay); the official source stays the OpenAI Astra announcement, as the page for the model used only. The author says the prompt is in a reply this library could not fetch, so the prompt is labelled as @LufzzLiz’s Chinese rendering. Boundaries: author claim plus a short video, no playable build, depth of play unverified; League of Legends IP belongs to Riot, so this records a model-capability showcase and distributes no game files; no invented numbers, unrated. Preview is the author’s video poster (champion select, Ahri, black band cropped). models.json also gains an optional relatedModelIds: the demo card and the GPT-6 Astra card point at each other, and detail pages get a “Related on this board” block rendering the full card. Models 2→3, dataset version 1.1.0.
站点：给本项目点个 GitHub StarSite: Star this project on GitHub
Hao 澄清「打星」指的是给 GitHub 项目 mostdesign01-sudo/ai-up-lab 点 Star，于是加一组站内推荐入口，全部指向公开仓库页。三处：站头导航右侧一枚紧凑的 GitHub「Star」药丸；首页左栏三库入口下方一张「给本项目点个 Star」小卡，写明仓库名与一句说明；页脚第一栏一行「觉得有用？去 GitHub 给本项目点个 Star」。Star 数由浏览器直接读 GitHub 公开 API 后填入，本机缓存一小时，读不到或为 0 就不显示，不编数字。这是给仓库点 Star，与卡片上的 ★ 编辑星级（JSON 里的 stars）和 ☆ 访客收藏（localStorage / /favorites/）是三回事，后两者原样保留。静态站，无密钥，只用公开 URL。Hao clarified that “stars” meant starring the GitHub project mostdesign01-sudo/ai-up-lab, so the site now carries a set of star prompts, all pointing at the public repo page. Three spots: a compact GitHub “Star” pill on the right of the header nav; a small “Star this project on GitHub” card in the home rail below the three library links, naming the repo with a one-line note; and one footer line, “Find this useful? Star the project on GitHub”. The star count is read in the browser from GitHub’s public API, cached locally for an hour, and simply stays hidden when unavailable or zero — no invented numbers. This stars the repo; it is separate from the ★ curator rating (stars in JSON) and the ☆ visitor favorites (localStorage / /favorites/), both of which stay as they were. Static site, no secrets, public URLs only.
按 Hao 要求「给网站加个打星」，分两层。一是编辑星级：cases / html-items / agent-ui / models 四套 JSON 新增可选 stars（1–5 整数），卡片与详情以 ★ 显示；是编辑对质量的判断，不是 GitHub star 数，不填就不显示。先给 12 条已精选条目打了 4–5 星（官方 Guides、隔夜销售外呼、CSS Zen Garden、scroll-driven animations、shadcn/ui 等），模型卡暂不打，厂商数字未核实。二是访客收藏：卡片右上角与详情页的 ☆ 点一下存进浏览器 localStorage，再点取消；新页 /favorites/「我的收藏」只读本机，导航 / 首页侧栏 / 页脚可进，可一键清空。静态站，无账号无后端，换设备不同步。At Hao’s request (“add stars to the site”), two layers. Curator stars: cases / html-items / agent-ui / models JSON gain an optional integer stars (1–5), shown as ★ on cards and detail pages. It is an editorial quality call, not GitHub stargazers, and unset means nothing is shown. Twelve already-featured items start at 4–5 (official Guides, overnight sales outbound, CSS Zen Garden, scroll-driven animations, shadcn/ui, …); model cards stay unrated since vendor figures are unverified. Visitor favorites: the ☆ on every card corner and detail page saves to the browser’s localStorage; tap again to remove. New /favorites/ (My favorites) reads local storage only, reachable from nav, home rail, and footer, with a clear-all button. Static site: no account, no backend, no cross-device sync.
Grok Bot：收录 Mobile MCP（AI 直接点 iOS / Android App）Grok Bot: Mobile MCP (agents tapping iOS / Android apps)
按 Hao 要求收录 @Ryrenz 9 月 8 日中文帖介绍的 Mobile MCP（mobile-next/mobile-mcp，Apache-2.0）。主条目指 GitHub README：MCP 服务器把 iOS / Android 模拟器、仿真器、真机接给 Claude Code / Codex / Gemini CLI / Copilot / Cursor 等 MCP 客户端，优先无障碍快照、退回截图坐标，工具覆盖装卸 App、点滑输入、录屏、logcat 与崩溃报告；默认 stdio，可 SSE + Bearer。当日核对 6468 star / 562 fork（帖内写 6353，已按实数）、最新 1.0.3。README 明写默认向 PostHog 发匿名遥测，MOBILEMCP_DISABLE_TELEMETRY=1 可关。配套 mobilewright / mobilecli 已在 README 核到，作次链接；Mobile Next Cloud 为付费云真机，本库未试。预览图取 GitHub 社交卡。归工程 / 编程 / 自动化，默认需审批。Grok Bot 96→97。At Hao’s request, catalogued Mobile MCP (mobile-next/mobile-mcp, Apache-2.0) from @Ryrenz’s 8 Sept Chinese post. Primary entry points to the GitHub README: an MCP server exposing iOS / Android simulators, emulators, and real devices to MCP clients such as Claude Code / Codex / Gemini CLI / Copilot / Cursor; accessibility snapshot first, screenshot-coordinate fallback; tools cover app install/remove, tap/swipe/type, screen recording, logcat and crash reports; stdio by default, SSE + Bearer optional. Checked that day: 6,468 stars / 562 forks (the post said 6,353; live count used), latest 1.0.3. README states anonymous PostHog telemetry is on by default; MOBILEMCP_DISABLE_TELEMETRY=1 turns it off. Companions mobilewright / mobilecli verified in the README and linked as secondary; Mobile Next Cloud is a paid device service we did not try. Preview is the GitHub social card. Filed under engineering / coding / automation, approval required by default. Grok Bot 96→97.
按 Hao 要求新增 /models/：追踪最新前沿模型与值得看的社区 showcase 的看板，与路径、组合同属实验室栏目，不是第四套内容库。首条 GPT-6 Astra：官方页链 OpenAI 公告与安全综述（2026-09-03，基准分 / 推出节奏 / 价格均为官方自述，未独立核实）；社区 showcase 链 @YouWareAI 9 月 8 日视频帖（「720+ 条作品」为 YouWare 自述；帖内 t.co 只指向视频，Astra Replay 页面未确认公开 URL，不写深链）。预览图取该视频封面。已在库的 Claude Fable 5.1 一并列为第二张卡，指回原 AI 资讯条目、不重复正文。数据在 data/models.json，导航 / 首页侧栏 / 页脚可进。At Hao’s request, added /models/: a board tracking latest frontier models and notable community showcases. Sits with Paths and Combos as a lab section — not a fourth library. First card is GPT-6 Astra: official links go to OpenAI’s announcement and safety overview (2026-09-03; benchmarks / rollout / pricing are vendor claims, unverified here); the community showcase links @YouWareAI’s 8 Sept video post (“720+ builds” is YouWare’s count; the post’s t.co resolves to the video and the Astra Replay page has no confirmed public URL, so no deep link). Preview is that video poster. The existing Claude Fable 5.1 entry is listed as a second card pointing back to the AI-news item without repeating copy. Data in data/models.json; reachable from nav, home rail, and footer.
HTML：收录 SceneAI（AETHER Claude 3Portals 演示同源）HTML: SceneAI (source of AETHER’s Claude 3Portals demo)
按 Hao 要求收录 @aether_uiux 的 Claude 网页设计演示片（3Portals：轻色 SaaS 首屏、玻璃拟态面板、紫色变形 3D 团块）。核对后帖子正文与自回复都没有可粘贴提示词，自回复链到 SceneAI（sceneai.art，Orixadesign 的付费 UI 提示库，140+ 提示 / 70+ 动态背景），故以 SceneAI 为主条目、原帖作次链接与证据。归 HTML 收集，与 MotionSites / MeiGen / Taste Skill 同属设计提示参考。3Portals 疑为虚构演示品牌，不写产品站。预览图取自视频封面。HTML 41→42。At Hao’s request, catalogued @aether_uiux’s Claude web-design clip (3Portals: light SaaS hero, glass panel, morphing purple 3D blob). On checking, neither the post nor the self-reply carries a pasteable prompt; the self-reply links to SceneAI (sceneai.art, Orixadesign’s paid UI prompt library, 140+ prompts / 70+ animated backgrounds), so SceneAI is the primary item with the post as secondary link and evidence. Filed in the HTML Collection alongside MotionSites / MeiGen / Taste Skill as a design-prompt reference. 3Portals appears to be a fictional demo brand, so no product URL. Preview taken from the video poster. HTML 41→42.
Grok Bot 增收 LaoZhang 安全第一天文档审阅教程、AI Maker 七人小队手机工作流。HTML 增收 HTML Presentation Editor（MIT 单文件）、Ptree 可交互生命树（HN 49559069）。Agent UI 增收 ThreeUI Community：Temple Night、Animated Top Dock。X @bot 网页检索无新的可复现帖。Grok Bot: LaoZhang safe first-day document review + AI Maker seven-Bot phone team. HTML: HTML Presentation Editor (MIT single-file) + Ptree interactive tree of life (HN 49559069). Agent UI: ThreeUI Community Temple Night + Animated Top Dock. No new reproducible @bot posts via web search.
Grok Bot 增收 AI Builder Club 实操指南、Gauraw 安全向完整指南、gawkbot 开源替代。HTML 增收 Laser Graffiti（HN 49489376）。Agent UI 无新增免费父组件。X @bot API 403，改走网页检索。Grok Bot: AI Builder Club practical guide, Gauraw control guide, gawkbot OSS alternative. HTML: Laser Graffiti (HN 49489376). Agent UI: no new free parent components. X @bot API 403; used web search instead.
资讯：收录 Claude Fable 5.1News: listed Claude Fable 5.1
收入 Anthropic Fable 5.1 / Mythos 5.1 官方公告为 AI 资讯（非 Grok Bot 用法）。主链公告，附官方 X 视频帖。官方数字未独立核实。Catalogued Anthropic’s Fable 5.1 / Mythos 5.1 announcement as AI news (not a Grok Bot playbook). Primary link is the post; official X video attached. Official figures unverified here.
9月2日 · 星期三 · 9 条9/2 · Wednesday · 9 items
站点：首页说清核验精选定位Site: homepage states curated-verified positioning
首页加短定位带：人核过的用法、路径、三库拣货。也看 grokbots.best（第三方安装目录，链到库内案例）。OG/meta 同步，不改海报和 logo。Homepage gained a short positioning band: human-checked use-cases, paths, three-library picks. Also see grokbots.best (third-party install directory, linked to the library case). OG/meta updated. No poster or logo rewrite.
站点：上线三库交叉组合 / CombosSite: Combos across the three libraries
新增 /combos/：四条精选配方，各取一条 HTML 壳、一件 Agent UI、一个 Grok Bot 案例。给全员课和 Design 工作室拣货，不是第四面墙。路径仍管玩法。Added /combos/: four curated recipes, each picking one HTML shell, one Agent UI piece, and one Grok Bot case. Kits for all-hands class and Design studio — not a fourth card wall. Paths stay the playbooks.
站点：广场本周可抄接到路径Site: plaza “this week’s steal” points at Paths
广场 digest 顶栏改成本周可抄：默认推「日更 → 一条可发草稿」，链到 /paths/daily-to-draft/。今日看点里对得上的增收会带相关路径。不改海报，也不做第四面墙。The plaza digest strip now features this week’s steal: default daily digest → one draft, linking to /paths/daily-to-draft/. Matching digest notes get a related-path hint. No poster rewrite, no fourth card wall.
路径详情加「开跑」Path detail: Run this path
路径详情加「开跑」：一键复制丢进对应桌的 brief。营销桌 / 内容桌 / 工程工头 / 设计味。只交草稿，站点不代发。Path pages gained 开跑 / Run this path: copy a paste-ready brief into the matching desk. Marketing / content / engineering / taste. Drafts only; the site does not send.
新增 /paths/：五条可点开玩法（内容桌、营销桌草稿停笔、工程工头、日更出草稿、设计味）。广场和导航可进。不堆第四面墙卡。Added /paths/: five clickable playbooks (content desk, marketing draft-only, engineering foreman, daily→draft, design taste). Reachable from plaza and nav. Not a fourth card wall.
增收 Lingxi Li 的 Grok Bot 工程组织长文（AYi 中文提纯为第二来源）。专科工头 + 状态机 + 人守闸门。规模数字未核实。Grok Bot 85→86。Added Lingxi Li’s Grok Bot engineering-org article (AYi’s Chinese digest as second source). Specialist bosses + state machine + human gates. Scale numbers unverified. Grok Bot 85→86.
Grok Bot：+2 awesome-grok-bot、How Do I Use AI 小队指南Grok Bot: +2 awesome-grok-bot and How Do I Use AI agent-team guide
增收 ZeroPointRepo 的 awesome-grok-bot（已核验 Skill/Plugin/MCP 目录，区别于 grokbots.best）与 How Do I Use AI 的无代码 agent-team 四周 rollout 指南。修正 meta.count 与真实条数对齐。Grok Bot 83→85。Added ZeroPointRepo’s awesome-grok-bot (verified skills/plugins/MCP directory, distinct from grokbots.best) and How Do I Use AI’s no-code agent-team four-week rollout guide. Fixed meta.count to match real length. Grok Bot 83→85.
增收 Hieu Dinh 的社区发现站 grokbots.best。第三方公开模板目录，不是官方名录。Grok Bot 82→83。Added Hieu Dinh’s community directory grokbots.best. Third-party public templates, not the official catalog. Grok Bot 82→83.
HTML：Refero Styles 补 Abraham 那条来源HTML: Refero Styles + Abraham John source
Abraham John（@Abmankendrick）8/31 又推了 Refero Styles（2000+ DESIGN.md）。条目已在库，只补第二来源，不另开卡。Abraham John (@Abmankendrick) reshared Refero Styles (2,000+ DESIGN.md) on 31 Aug. Item already in the library; added as a second source, no new card.
Grok Bot：增收 figma broGrok Bot: + figma bro
增收 John Bai 公开的 Figma 设计系统搭档模板。走 Figma 连接器改文件。Grok Bot 81→82。Added John Bai’s public Figma design-system partner template. Edits through the Figma connector. Grok Bot 81→82.
增收 Lonely 转发的非官方教程：用 Tailscale 把 Grok Bot 电脑和自己电脑拉进同一张网再 SSH。标了不要暴露 22 端口。Grok Bot 78→79。Added Lonely’s share of an unofficial tutorial: join Grok Bot’s computer and your laptop on Tailscale, then SSH. Notes not to expose port 22. Grok Bot 78→79.
增收 B2TF 与无聊目录玩法Added B2TF and a boring-directory playbook
HTML 收集增收 B2TF 复古 iOS 应用目录（@jackcring）。Grok Bot 增收 robert. 的「无聊目录每晚长一页」社区玩法（联盟佣金未核实）。HTML 32→33，Grok Bot 77→78。HTML: B2TF vintage iOS app directory (@jackcring). Grok Bot: robert.’s “grow a boring directory a page a night” community playbook (affiliate commissions unverified). HTML 32→33, Grok Bot 77→78.
增收 OpenDesign：开源 Claude Design 替代，本机优先桌面应用，导出 HTML / PDF。来源 @sunmer575399。HTML 收集 25→26。Added OpenDesign, the open-source Claude Design alternative — local-first desktop app, HTML/PDF export. Source: @sunmer575399. HTML 收集 25→26.
Grok Bot：增收 2 条社区案例Grok Bot: +2 community cases
增收 六席 Bot 跑手游工作室：Rank'em、非官方 Linux 版 Grok Bot：从 Windows 安装包移植。来源分别为 Ryan Perry 案例研究（兼 HN item 49447324）与 jakob-bu 非官方 Linux 移植（HN item 49467702）。Grok Bot 72→74。Added Six-bot mobile game studio: Rank'em; Unofficial Linux Grok Bot (Windows installer port). Sources: Ryan Perry’s case study (also HN item 49447324) and jakob-bu’s unofficial Linux port (HN item 49467702). Grok Bot 72→74.
HTML 收集：增收 GlassBox 与 Voronoi GoHTML Collection: add GlassBox and Voronoi Go
从 Hacker News 增收 GlassBox：浏览器指纹台、Voronoi Go：没有网格的围棋（item 49421948、item 49468816）。均带 3:2 自托管预览。HTML 收集 23→25。From Hacker News: GlassBox: browser fingerprint bench; Voronoi Go: Go without the grid (item 49421948、item 49468816). Both have self-hosted 3:2 previews. HTML 收集 23→25.
来源 https://x.com/grgerwcwetwet/status/2092937412490260513 ：五个给 Agent 抄作业的 UI 站。Agent UI 20→25。Transitions.dev 只收免费区。From https://x.com/grgerwcwetwet/status/2092937412490260513: five UI kits for Agents to copy from. Agent UI 20→25. Transitions.dev free gallery only.
8月27日 · 星期四 · 6 条8/27 · Thursday · 6 items
Grok Bot：增收 Matt Palmer 入门视频Grok Bot: add Matt Palmer’s Getting Started video
来源 https://x.com/mattyp/status/2092286281266962681 ：Getting Started with Grok Bot，含 X→Notion、个人软件、Cursor 编码、搜索四条工作流。Grok Bot 71→72。From https://x.com/mattyp/status/2092286281266962681: Getting Started with Grok Bot, including X→Notion, personal software, Cursor coding, and search. Grok Bot 71→72.
来源 X：@sucodeee 的 MetalForge 进度条 shader（https://metalforge.xyz/），以及风间推荐的 Buttermax 交互作品集（https://buttermax.net/）。Agent UI 18→20。From X: @sucodeee’s MetalForge progress-bar shaders (https://metalforge.xyz/), and Buttermax, the interaction portfolio site recommended by 风间 (https://buttermax.net/). Agent UI 18→20.
Grok Bot：补录今早漏掉的 6 个官方目录角色Grok Bot: +6 official catalog roles missed in the morning ingest
来源 https://x.ai/bot/use-cases ：Daily Briefing Writer、Executive Assistant、Newsletter Writer、Prospecting Plan Builder、Status Report Writer、CRM Operations Manager。今早 36 角色入库（29→65）时被当成近重复跳过，现补齐 65→71。From https://x.ai/bot/use-cases: Daily Briefing Writer, Executive Assistant, Newsletter Writer, Prospecting Plan Builder, Status Report Writer, CRM Operations Manager. These were skipped as near-dupes during the morning 36-role ingest (29→65); catalog is now 65→71.
Grok Bot：官方用法目录增收 36 个角色Grok Bot: +36 roles from the official use-cases catalog
来源 https://x.ai/bot/use-cases （及扩容帖退款追回经理）：公寓侦察、差旅协调、日历协调、招聘筛选、入职经理、客户研究、会前准备、销售通话教练、交易台、续约台、幻灯片现场更新、Beta 采用、通话 FAQ、文档审计、需求溯源、产品反馈、玩法测试、原型构建、云端 Agent 编排、安全问卷、供应商门户、合同台、退款追回、SEO/AEO、社媒、付费素材策略、LinkedIn 投放、营销日历、周边履约、社区运营、关键事件监测、活动嘉宾筛选、赋能交付、工单分拣、演示文稿设计、对内传播。已有 29 条对应角色未重复收录。@bot 时间线本次未能读取。From https://x.ai/bot/use-cases (plus Refunds manager on the expansion post): apartment scout, travel and calendar coordinators, hiring screener, onboarding, account research, meeting prep, sales call coach, deal desk, renewal desk, live deck updater, beta adoption, call FAQ miner, docs auditor, feature-request tracker, product feedback, playtest, prototype builder, cloud-agent orchestrator, security questionnaire, vendor portal, contract desk, refunds claims, SEO/AEO, social, paid-creative strategy, LinkedIn campaigns, marketing calendar, merch fulfillment, community ops, compelling-events monitor, event guest screener, enablement fulfillment, ticket triage, presentation designer, internal comms. The existing 29 matching roles were not duplicated. @bot timeline was not readable this run.
8月26日 · 星期三 · 2 条8/26 · Wednesday · 2 items
HTML 收集：从 Hacker News 增收 Don't Wordle 与 RACK-02HTML Collection: Don't Wordle and RACK-02 from Hacker News
来源 Hacker News（news.ycombinator.com / @hackernews）：Don't Wordle（item 49432319，每日 HTML 字谜）与 RACK-02（item 49419351，单文件 techno 机）。均带 3:2 自托管预览。Grok Bot 与 Agent UI 今日无新增。From Hacker News (news.ycombinator.com / @hackernews): Don't Wordle (item 49432319, daily HTML word game) and RACK-02 (item 49419351, techno machine in one HTML file). Both have self-hosted 3:2 previews. No Grok Bot or Agent UI adds today.
三库更新：Grok Bot +4，HTML +1，Agent UI +4All three libraries: Grok Bot +4, HTML +1, Agent UI +4
Grok Bot：官方扩容帖补游戏美术、办公室经理、客服政策内退款，另收 Debbie O'Brien 手机语音买菜教程。HTML 收集：slideVars（3:2 自托管预览）。Agent UI：ThreeUI Community 新增 Koi Studies、Bestsellers Book Showcase、Country Towers、Character Carousel（均免费、免登录，带预览）。Grok Bot: remaining official expansion roles Game Artist, Office Manager, Customer Support in-policy refunds, plus Debbie O'Brien’s mobile voice grocery tutorial. HTML Collection: slideVars (self-hosted 3:2 preview). Agent UI: ThreeUI Community adds Koi Studies, Bestsellers Book Showcase, Country Towers, Character Carousel (free, no login, with previews).
8月25日 · 星期二 · 2 条8/25 · Tuesday · 2 items
Grok Bot：收录官方扩容帖 3 个新角色Grok Bot: three new roles from the official expansion post
来源 https://x.ai/news/grok-bot-more-plans ：数字断舍离、会议替身、建站 Bot。HTML 收集与 Agent UI 今日无新增（避免无预览的色块条目）。Source https://x.ai/news/grok-bot-more-plans : Digital Declutterer, Meeting Stand-in, Website Builder. No HTML Collection or Agent UI adds today (avoid color-block pins without previews).
Grok Bot 收录 Kanekoa 视频研究剪辑案例Grok Bot lists the Kanekoa video-research cutdown case
社区案例：@KanekoaTheGreat 用 Grok Bot 将 38 条 Seheult 相关视频转写、精选 13 段剪成 19:35，并生成带时间戳的 Markdown。卡片使用推文截图作为预览。Community case: @KanekoaTheGreat used Grok Bot to transcribe 38 Seheult-related videos, select 13 segments into a 19:35 cut, and generate timestamped Markdown. The card uses a tweet screenshot as the preview.
8月24日 · 星期一 · 6 条8/24 · Monday · 6 items
补全预览缩略图Fill in preview thumbnails
HTML 收集与 Agent UI 的 masonry 卡片改为自托管预览缩略图（public/previews），Pinterest 墙显示真实界面/效果，而不是纯色海报。图片路径走 Astro base，适配 GitHub Pages。缺图或加载失败时仍回退库海报。同时把 ThreeUI 原文改成现用的 /{category}/{slug} Community 路由。HTML Collection and Agent UI masonry cards now use self-hosted preview thumbnails (public/previews). The Pinterest wall shows real UI/effects instead of flat-color posters. Image paths go through the Astro base for GitHub Pages. Missing or failed images still fall back to the library poster. ThreeUI source links now use the current /{category}/{slug} Community routes.
界面：三张海报 + Pinterest 式浏览墙UI: three posters + a Pinterest-style wall
首页改为三张库海报（Grok Bot / HTML 收集 / Agent UI）加跨库精选瀑布流。Grok Bot 总览移到 /cases/。各库与精选/最新/类型页改为视觉优先的 masonry 卡片：有 previewImage 用图，否则用库海报，不再用时间线当主界面。Home is now three library posters (Grok Bot / HTML Collection / Agent UI) plus a cross-library featured masonry. Grok Bot overview moved to /cases/. Each library and featured/latest/types page is now visual-first masonry: previewImage when present, otherwise the library poster. The timeline is no longer the main UI.
界面：字号、密度与表面打磨UI: type size, density, and surface polish
全站改用 Geist + Noto Sans SC（中文一等），收紧标题字重与字距；卡片、顶栏分段开关、库选择、时间线与搜索更密一档，边框与阴影更克制。三库配色（琥珀 / 青绿 / 靛紫）与案例、HTML、Agent UI 数据不变。The site now uses Geist + Noto Sans SC (Chinese first), with tighter title weight and tracking. Cards, header segmented controls, library picker, timeline, and search are one step denser; borders and shadows are quieter. The three library colors (amber / teal / violet) and the Grok, HTML, and Agent UI data are unchanged.
站点更名为 AI UP LAB，并新增 Agent UI 库Site renamed AI UP LAB, Agent UI library added
对外品牌定为 AI UP LAB：可复用的 AI 精选实验室 · 案例 / 页面 / Agent UI。顶栏现为三库：Grok Bot（琥珀）、HTML 收集（青绿）、Agent UI（靛紫）。Agent UI 首批收录 ThreeUI Community 免费组件（MIT，免登录），可复制 prompt / Skill.md 换主题；不含 Pro，也不提供 MCP。Grok Bot 与 HTML 数据、路由保持不变。Public brand is AI UP LAB: a reusable AI collection lab · cases / pages / Agent UI. The header now has three libraries: Grok Bot (amber), HTML Collection (teal), Agent UI (violet). Agent UI’s first batch is free ThreeUI Community components (MIT, no login); copy a prompt / Skill.md to retheme. No Pro, no MCP. Grok Bot and HTML data and routes stay the same.
界面：近日更新时间线与收录理由UI: recent-update timeline and why we listed it
双库首页增加按 updatedAt 分组的「今日收录 / 近日更新」，卡片展示收录理由（qualityNote），并加上当前库搜索、上手路径与 /search 两库检索。不是行业新闻站。The dual-library home gained “listed today / recent updates” grouped by updatedAt. Cards show why we listed it (qualityNote), plus in-library search, a start path, and /search across both libraries. Not an industry news site.
Bot 补 1 条官方 Skill 教学，HTML 收集 +5Bot adds 1 official Skill lesson; HTML Collection +5
Grok Bot：新增官方「Teach a task → Skill → Routine」演示教学案例。HTML 收集：新增 CSS Zen Garden、Scroll-driven Animations、Squoosh、Clippy、css-doodle。Grok Bot: added the official “Teach a task → Skill → Routine” demo case. HTML Collection: added CSS Zen Garden, Scroll-driven Animations, Squoosh, Clippy, and css-doodle.
8月23日 · 星期日 · 2 条8/23 · Sunday · 2 items
v1.0.0 种子数据入库v1.0.0 seed data checked in
收录 20 条公开资料中的精选用法，覆盖官方文档与发布稿、社区多 Agent 实践，以及 Cursor Automations 相邻的工程工厂模式。原始记录见 data/cases.json。Listed 20 selected uses from public sources, covering official docs and launch posts, community multi-agent practice, and Cursor Automations-adjacent engineering-factory patterns. Raw records are in data/cases.json.
新增独立「HTML 收集」库Added a separate HTML Collection library
顶部增加 HTML 收集标签，数据在 data/html-items.json，与 Grok Bot 案例分开。首批收录 12 条可公开访问的演示、模板、组件与工具页。The header gained an HTML Collection tab. Data lives in data/html-items.json, separate from Grok Bot cases. First batch: 12 publicly reachable demos, templates, components, and tool pages.
