AI Browser v6.1.0

AI 浏览器 v6.1.0

本版本引入拓展框架：全部内置功能均改为随应用提供的第一方拓展，桌面版与 iPad 版使用同一拓展格式。安装包不再包含模型文件，大小约为上一版本的一半。

新增

• 拓展框架（框架 2 级）：解题、拍题、GeoGebra、物理、百科、词典、代码与稍后读均为预装的 aib-feature/1 拓展，可在“设置 → 拓展”中停用或重新启用。停用后，对应工具从工具栏与主页中移除；重新启用后恢复到原位置。
• 贡献点：拓展可提供工具面板项、主页入口、回答内容块（如 graph、word）、文件导入、设置页、URL 协议与原生视图。用户拓展可提供自带图标的工具（图标经过安全过滤）；URL 协议、原生视图与内置命令仅限第一方拓展使用。
• 开发文档内置于应用：“设置 → 拓展 → 开发文档”（aibrowser://docs/extensions）提供开发指南、贡献点参考与迁移记录。拓展卡片列出该拓展提供的内容及适用平台。
• 以本应用打开词典文件：在“访达”中可通过“打开方式”以本应用打开 .mdx 与 .mdd 文件，文件将由词典拓展导入。

变更

• 安装包不含模型文件：知识库所用的嵌入模型 EmbeddingGemma 300M 改为在首次使用知识库时下载到用户数据目录（支持断点续传与镜像下载源），与 HY-MT 翻译模型的处理方式一致。AI-Browser-mac-arm64.zip 由约 628 MB 减小到约 314 MB。

修复

• 日志：从终端启动后关闭终端等情况下，标准输出关闭后的写入失败不再引发无休止的错误记录与 CPU 占用，也不再被误判为崩溃。

测试

• 新增拓展框架测试：全部内置拓展加载、工具栏与主页入口顺序、每个工具均可打开、停用与恢复、回答内容块与文件导入、用户拓展的图标过滤与越权清单拒绝、开发文档显示。
• 新增嵌入模型首次使用测试：模拟下载源提供真实模型文件，确认首次添加知识库内容时自动下载并能正常检索；打包脚本检查应用内不含任何模型文件。
• 新增日志安全测试：标准输出关闭后持续输出与未捕获异常均不引发循环，日志大小受限。

下载

文件	说明
AI-Browser-mac-arm64.zip	未签名的 AI Browser.app（macOS arm64）；模型在首次使用时下载
install.command / install.sh	解压 → 复制到“应用程序” → xattr -cr → 启动
AI-Browser-source.zip	源代码
README.md / README-zh.md	英文 / 中文说明

---

AI Browser v6.1.0

This release introduces the extension framework: every built-in feature is now a first-party extension shipped with the app, and the desktop and iPad versions use the same extension format. The download no longer contains model files and is about half the size of the previous release.

New

• Extension framework (level 2): Solve, Photo solve, GeoGebra, Physics, Wiki, Dictionary, Code and Read later are pre-installed aib-feature/1 extensions that can be turned off and on again in Settings → Extensions. When an extension is turned off, its tool is removed from the toolbar and the home page; turning it back on restores it in its original place.
• Contribution points: extensions can add tool-panel items, home-page entries, answer blocks (such as graph and word), file importers, settings pages, URL schemes and native views. User extensions can add tools with their own icons (icons are sanitised); URL schemes, native views and built-in commands are reserved for first-party extensions.
• Developer documentation built in: Settings → Extensions → Developer Docs (aibrowser://docs/extensions) contains the developer guide, the contribution point reference and the migration record. Each extension card lists what the extension provides and the platforms it supports.
• Open dictionary files with the app: .mdx and .mdd files can be opened with the app from Finder's "Open With" menu; the dictionary extension imports them.

Changes

• No model files in the download: the knowledge base's embedding model, EmbeddingGemma 300M, is now downloaded into the user data folder the first time the knowledge base is used (resumable, with mirror sources), in the same way as the HY-MT translation model. AI-Browser-mac-arm64.zip shrinks from about 628 MB to about 314 MB.

Fixes

• Logging: when standard output closes (for example after closing the terminal the app was started from), failed writes no longer cause endless error logging and CPU use, and are no longer treated as a crash.

Testing

• New extension framework tests: all built-in extensions load, toolbar and home-page order, every tool opens, turning extensions off and on, answer blocks and file importers, user extension icon sanitising and refusal of over-privileged manifests, and the developer documentation.
• New embedding model first-use test: a mock download source serves the real model file, and the first knowledge-base addition downloads it automatically and search works; the packaging script checks that the app contains no model files.
• New logging safety tests: continued output and uncaught errors after standard output closes cause no loop, and log size is capped.

Downloads

File	Description
AI-Browser-mac-arm64.zip	Unsigned AI Browser.app (macOS arm64); models download on first use
install.command / install.sh	Unzip → copy to Applications → xattr -cr → launch
AI-Browser-source.zip	Source code
README.md / README-zh.md	English / Chinese documentation

Created by Grok

AI Browser v6.0.1

AI 浏览器 v6.0.1

本版本为 v6.0.0 的修复更新，修复了打开界面后 CPU 与 GPU 持续满载的问题，并包含与 iPad 3.0.1 同步的多项修复。

修复

• CPU 与 GPU 持续满载：当本机正在运行本地模型框架（如 Ollama、LM Studio）时，打开主页即会导致 CPU 与 GPU 占用持续接近 100%。原因是主页的模型选择器每次加载都会触发“模型已更新”通知，该通知又使选择器重新加载，形成无休止的循环。现仅在本地框架的状态或模型列表实际发生变化时才发出通知，空闲时 CPU 占用恢复到正常水平。
• 词典显示区域：词条区域按内容自动调整高度，并在词典自带的脚本、字体与图片加载完成后重新测量，释义与例句不再被截断；词典自带的样式表、脚本与字体按原样加载，entry://、sound:// 与 @@@LINK= 均可使用；缺少资源文件时给出提示。
• 词典 .mdd 资源文件：导入时可识别同名 .mdd 与分卷文件（如 .1.mdd），不受文件名大小写、空格及 Unicode 组合形式（NFC/NFD）差异影响；词典设置中显示已关联的 .mdd，并可通过“添加资源…”在之后补充。
• 刷新：工具标签页刷新后停留在原结果并重新执行查询。
• GeoGebra 绘图：绘图区默认约 55% 宽度、坐标轴 1:1，新增可拖动分隔条与“收起回答 / 展开回答”按钮，状态会被记住。
• Linux 桌面（虚拟机）：在触控屏上以触控板方式操作（单指移动指针、轻点左键、长按右键、双指滚动、轻点后拖动），鼠标操作不变。

测试

• 新增空闲性能测试：在本地框架运行时打开主页，断言全部进程空闲 CPU 低于 15%，页面无持续布局、样式计算或 requestAnimationFrame 循环，且不会反复探测本地框架。
• 新增词典、.mdd 识别、刷新、GeoGebra 分栏与虚拟机触控操作的自动化测试。

下载

文件	说明
AI-Browser-mac-arm64.zip	未签名的 AI Browser.app（macOS arm64）
install.command / install.sh	解压 → 复制到“应用程序” → xattr -cr → 启动
AI-Browser-source.zip	源代码
README.md / README-zh.md	英文 / 中文说明

---

AI Browser v6.0.1

This is a bug-fix update to v6.0.0. It fixes CPU and GPU usage staying at full load after the window opens, and includes the fixes shared with iPad 3.0.1.

Fixes

• CPU and GPU at full load: while a local model framework (such as Ollama or LM Studio) was running on the Mac, opening the home page kept CPU and GPU usage near 100%. Each time the home page's model picker loaded, it sent a "models changed" notification, which made the picker load again, in an endless loop. The notification is now sent only when the local frameworks' status or model list actually changes, and idle CPU usage is back to normal.
• Dictionary display area: entries size themselves to their content and are measured again after the dictionary's own scripts, fonts and images load, so definitions and examples are no longer cut off. The dictionary's own style sheets, scripts and fonts are loaded as provided, entry://, sound:// and @@@LINK= work, and a note appears when resource files are missing.
• Dictionary .mdd resource files: import recognises the .mdd with the same name and split files (such as .1.mdd) regardless of letter case, spaces and Unicode normalisation (NFC/NFD). Dictionary settings list the linked .mdd files, and "Add resources…" adds more later.
• Refresh: reloading a tool tab stays on the result and runs the query again.
• GeoGebra plots: the plot takes about 55% of the width by default with 1:1 axes, with a draggable divider and a "Hide answer / Show answer" button; both settings are remembered.
• Linux desktop (virtual machine): on touch screens it works like a trackpad (one finger moves the pointer, tap to left-click, long press to right-click, two fingers to scroll, tap-then-drag to drag); mouse input is unchanged.

Testing

• New idle performance test: with a local framework running, opening the home page must keep idle CPU across all processes below 15%, with no continuous layouts, style recalculations or requestAnimationFrame loops, and no repeated probing of local frameworks.
• New automated tests for the dictionary, .mdd recognition, refresh, the GeoGebra split layout and virtual machine touch input.

Downloads

File	Description
AI-Browser-mac-arm64.zip	Unsigned AI Browser.app (macOS arm64)
install.command / install.sh	Unzip → copy to Applications → xattr -cr → launch
AI-Browser-source.zip	Source code
README.md / README-zh.md	English / Chinese documentation

Created by Grok

AI Browser v6.0.0

AI 浏览器 v6.0.0

本版本是一次大版本更新：对话改为会话形式并可随时停止生成，回答采用统一的结构化格式，新增文档排版引擎与多种生成助手，Task 模式可使用全部浏览器工具，模型下载更快且新增魔搭（ModelScope）下载源。

对话

• 会话：追问显示在上一条回答下方，形成一段连续的对话；对话随会话保存，重启应用后仍可继续。较早的轮次在超出模型上下文时自动精简，Task 模式同样适用。
• 对话列表：新增对话列表（菜单“对话记录”），可搜索、重新打开、重命名与删除。
• 停止生成：回答过程中发送按钮变为停止按钮（亦可按 Esc 或 ⌘.），可同时终止回答、正在运行的工具与 Agent 步骤；已生成的内容予以保留并标注“已中止”。
• 思考过程：模型的思考内容在生成时实时显示，并默认折叠。

回答格式

• 回答统一采用标题、列表、表格、引用与 LaTeX 公式，排版更加清晰。
• 化学方程式：以 mhchem 记法排版，并自动检查原子守恒与电荷守恒，未配平时给出提示。
• 函数图像：回答中的函数可直接在 GeoGebra 中作图。
• 单词卡片：值得记忆的英文单词以卡片形式列出，点击即可查看词典释义。

文档与排版

• 新增排版引擎：支持 A4、A5、A3、B5、Letter、幻灯片、卡片等页面尺寸；流式与绝对定位元素、旋转、锚点、自动分页与重叠检查；样式采用预设的字号、颜色、间距与主题。
• 新增生成助手：文章与讲义、幻灯片、表格与图表（柱状图、条形图、折线图、饼图、散点图）、试卷与错题本（答案可不显示、随题显示或附于文末）、单词卡片、思维导图。
• 每份文档在对话中以卡片形式显示缩略图，可打开预览，并导出 PDF 或 HTML；导出的文件同时出现在下载列表中。
• AI 在排版过程中会收到页面预览图，用于检查并调整版面。

Task 模式

• 所有工具统一登记，Task 模式下助手可使用全部浏览器功能：标签页、工具视图、分屏、空间、稍后阅读、页面助手、截图、外观设置、扫码等。
• 每一步以工具卡片显示，工具名称为中文，参数以简明方式呈现。

模型下载

• 多连接分段下载：大文件分为多个分段并行下载（默认 8 个连接），各分段均可续传，并自动检测停滞、重新分配剩余部分与切换镜像。
• 魔搭（ModelScope）：新增为下载源。默认自动模式会实测各下载源的速度并排序，失败时自动切换。下载源与连接数可在设置中调整。

GeoGebra

• 修复在 GeoGebra 尚未加载完成时发起作图导致长时间无响应的问题。页面加载（30 秒）、组件加载（60 秒，含重试）与作图调用（100 秒）均设有超时，并给出明确的中文提示。

测试

• 新增对话会话、分段下载、Task 模式工具注册、排版引擎与生成助手、答案格式等自动化测试。

下载

文件	说明
AI-Browser-mac-arm64.zip	未签名的 AI Browser.app（macOS arm64）
install.command / install.sh	解压 → 复制到“应用程序” → xattr -cr → 启动
AI-Browser-source.zip	源代码
README.md / README-zh.md	英文 / 中文说明

---

AI Browser v6.0.0

This is a major release. Conversations now continue as threads and can be stopped at any time, answers follow one structured format, a document layout engine with several generator assistants is added, task mode can use every browser tool, and model downloads are faster, with ModelScope as a new source.

Conversations

• Threads: follow-ups appear below the previous answer as one continuous conversation. Conversations are saved and can be continued after the app restarts. Earlier turns are trimmed automatically when they exceed the model's context, in task mode as well.
• Conversation list: a new conversation list (menu item 对话记录, Chats) lets you search, reopen, rename and delete conversations.
• Stop: while an answer is being generated, the send button becomes a Stop button (Esc or ⌘. also work). It ends the answer, any running tools and agent steps together, and keeps the partial answer marked as stopped.
• Thinking: the model's thinking is shown live and collapsed by default.

Answer format

• Answers consistently use headings, lists, tables, quotations and LaTeX formulas.
• Chemical equations are typeset in mhchem notation and checked for conservation of atoms and charge; unbalanced equations are flagged.
• Function graphs: functions in an answer can be plotted in GeoGebra directly.
• Word cards: English words worth learning are listed as cards; tapping a card opens the dictionary entry.

Documents and layout

• New layout engine: page sizes including A4, A5, A3, B5, Letter, slides and cards; flowing and absolutely positioned elements, rotation, anchors, automatic page breaks and overlap checks; styles use preset sizes, colors, spacing and themes.
• New generator assistants: articles and handouts, slides, tables and charts (column, bar, line, pie and scatter), exam papers and mistake notebooks (answers hidden, shown with each question, or at the end), word cards and mind maps.
• Each document appears in the conversation as a card with a thumbnail. It can be previewed and exported as PDF or HTML, and exported files also appear in Downloads.
• During layout the AI receives page previews to check and adjust the pages.

Task mode

• All tools are registered in one place, so in task mode the assistant can use every browser feature: tabs, tool views, split view, spaces, read later, the page assistant, screenshots, appearance settings, QR scanning and more.
• Each step is shown as a tool card with the tool's Chinese name and a short summary of its arguments.

Model downloads

• Multi-connection downloads: large files are downloaded in parallel segments (8 connections by default), each resumable, with automatic stall detection, reassignment of the remaining parts and mirror rotation.
• ModelScope is a new download source. In the default automatic mode the sources are ranked by measured speed, with automatic fallback. The source and the number of connections can be changed in Settings.

GeoGebra

• Fixed long unresponsive waits when a graph was requested before GeoGebra had finished loading. Page loading (30 s), applet loading (60 s, with a retry) and graph calls (100 s) now have time limits with clear messages.

Tests

• New automated tests cover conversation threads, segmented downloads, the task-mode tool registry, the layout engine and generator assistants, and the answer format.

Downloads

File	Description
AI-Browser-mac-arm64.zip	Unsigned AI Browser.app (macOS arm64)
install.command / install.sh	Unzip → copy to Applications → xattr -cr → launch
AI-Browser-source.zip	Source code
README.md / README-zh.md	English / Chinese README

Created by Grok

AI Browser iPad 3.1.0

AI Browser iPad 3.1.0 (build 310)

本版本引入拓展框架：全部内置功能均改为随应用提供的第一方拓展，iPad 版与桌面版使用同一拓展格式。

新增

• 拓展框架（框架 2 级）：解题、拍题、GeoGebra、物理、百科、词典、代码、稍后读与扫码均为预装的 aib-feature/1 拓展，可在“设置 → 拓展”中停用或重新启用。停用后，对应工具从工具栏“工具”面板与主页中移除；重新启用后恢复到原位置。
• 贡献点：拓展可提供工具面板项、主页入口、回答内容块（如 graph、word）、文件导入（文件类型与 UTType）、设置页、URL 协议与原生视图。用户拓展可提供自带图标的工具（图标经过安全过滤）；URL 协议、原生视图与内置命令仅限第一方拓展使用。
• 开发文档内置于应用：“设置 → 拓展 → 开发文档”（aibrowser://docs/extensions）提供开发指南、贡献点参考与迁移记录。拓展卡片列出该拓展提供的内容及适用平台。

修复

• 设置 → 关于：现显示 iPad 版本号与运行环境（iPadOS、WebKit、Node.js），不再显示桌面版的 Electron 与 Chromium 信息。
• 日志：标准输出关闭后，日志写入失败不再引发无休止的重复记录；node.log 每个文件上限为 5 MB，仅保留当前文件与一份旧文件，日志不会持续占用存储空间。

测试

• 新增拓展框架测试（桌面版与 iPad 版共用）：全部内置拓展加载、工具面板与主页入口顺序、每个工具均可打开、停用与恢复、回答内容块与文件导入、用户拓展的图标过滤与越权清单拒绝、开发文档显示，以及“关于”页面版本号。
• 新增日志安全测试：标准输出关闭后持续输出与未捕获异常均不引发循环，日志大小受限，测试工具在退出或中断时结束全部派生进程。
• 界面自动化测试在 iPadOS 18 与 iPadOS 26 模拟器上通过。

安装

下载 AI-Browser-iPad-unsigned.ipa，使用 AltStore、Sideloadly 等工具以自己的 Apple ID 侧载安装（iPadOS 16 或更高版本）。

---

AI Browser for iPad 3.1.0 (build 310)

This release introduces the extension framework: every built-in feature is now a first-party extension shipped with the app, and the iPad and desktop versions use the same extension format.

New

• Extension framework (level 2): Solve, Photo solve, GeoGebra, Physics, Wiki, Dictionary, Code, Read later and Scan are pre-installed aib-feature/1 extensions that can be turned off and on again in Settings → Extensions. When an extension is turned off, its tool is removed from the toolbar's Tools panel and the home page; turning it back on restores it in its original place.
• Contribution points: extensions can add tool-panel items, home-page entries, answer blocks (such as graph and word), file importers (file types and UTTypes), settings pages, URL schemes and native views. User extensions can add tools with their own icons (icons are sanitised); URL schemes, native views and built-in commands are reserved for first-party extensions.
• Developer documentation built in: Settings → Extensions → Developer Docs (aibrowser://docs/extensions) contains the developer guide, the contribution point reference and the migration record. Each extension card lists what the extension provides and the platforms it supports.

Fixes

• Settings → About: now shows the iPad version and runtime (iPadOS, WebKit, Node.js) instead of the desktop Electron and Chromium details.
• Logging: a failed log write after standard output closes no longer causes endless repeated logging; each node.log file is capped at 5 MB and only the current file and one older file are kept, so logs no longer keep consuming storage.

Testing

• New extension framework tests (shared by the desktop and iPad versions): all built-in extensions load, tool-panel and home-page order, every tool opens, turning extensions off and on, answer blocks and file importers, user extension icon sanitising and refusal of over-privileged manifests, the developer documentation, and the version on the About page.
• New logging safety tests: continued output and uncaught errors after standard output closes cause no loop, log size is capped, and the test tools end every spawned process when they exit or are interrupted.
• UI automation tests pass on the iPadOS 18 and iPadOS 26 simulators.

Installation

Download AI-Browser-iPad-unsigned.ipa and sideload it with your own Apple ID using AltStore, Sideloadly or a similar tool (iPadOS 16 or later).

Created by Grok

AI Browser iPad 3.0.1

AI Browser iPad 3.0.1 (build 301)

本版本为 iPad 3.0.0 的修复与改进更新。

新增

• 内置中文文档字体：思源宋体（Noto Serif SC，常规与粗体）与霞鹜文楷（LXGW WenKai），用于文档排版中的宋体与楷体文字；字体按 SIL 开源字体许可证（OFL 1.1）随应用提供。
• “设置 → 关于”显示 iPad 版本号。

修复

• 工具栏“工具”玻璃面板：轻点面板中的各项工具（解题、拍题、GeoGebra、物理、百科、词典、代码、稍后读、扫码）现可正常打开对应工具；“展开到整个区域”按钮同样可用。旁白（VoiceOver）与全键盘控制亦可激活各项工具。
• 选择工具后可立即再次打开“工具”面板，无需等待所选工具加载完成。
• 连续导入时，在上一个“文件”选择器尚未完全关闭时发起的下一次导入不再被忽略。
• 本机运行本地模型服务时，主页不再持续占用大量 CPU。
• 主页输入框下方的工具图标行在 iPad 上不再显示，所有工具统一从工具栏的“工具”面板打开。
• 词典显示区域：词条区域现按内容自动调整高度，并在词典自带的脚本、字体与图片加载完成后重新测量，释义与例句不再被截断。词典自带的样式表、脚本与字体按原样加载，entry:// 词条链接、sound:// 发音与 @@@LINK= 跳转均可使用；词条背景保持白色。缺少资源文件时，词条上方会给出提示。
• 词典 .mdd 资源文件识别：“文件”选择器可直接选择 .mdx 与 .mdd 文件；导入文件夹时可识别与词典同名的 .mdd 以及分卷文件（如 .1.mdd），文件名大小写、空格及 Unicode 组合形式（NFC/NFD）的差异不再影响匹配；文件夹中仅有一部词典时，其中的全部 .mdd 均归入该词典。词典设置中显示每部词典已关联的 .mdd 文件，并可通过“添加资源…”在之后补充 .mdd；单独选择 .mdd 时会自动归入同名词典。
• 刷新：在工具标签页中刷新后，页面停留在原结果并重新执行查询，不再回到工具首页。
• Linux 桌面（虚拟机）触控操作改为触控板方式：单指拖动相对移动指针，页面不再滚动或回弹；轻点为左键单击，长按为右键单击，轻点两下为双击，双指拖动为滚轮滚动，轻点后按住拖动为按住左键拖动。外接鼠标、触控板与 Apple Pencil 仍可正常使用。
• 对话中的 GeoGebra 绘图：绘图区默认占约 55% 宽度，坐标轴默认 1:1，并在布局变化或旋转屏幕时重新调整画布大小；绘图区与回答之间新增可拖动的分隔条（双击恢复默认比例）；新增“收起回答 / 展开回答”按钮，收起状态与分隔比例会被记住。

测试

• 新增界面自动化测试：以真实点按依次打开面板中的每一项工具并确认工具已打开，同时测试展开按钮；在 iPadOS 18 与 iPadOS 26 模拟器上运行。
• 新增测试：仿 OALD 结构的词典样例（自带样式表、脚本、字体、发音与跳转）的显示与自动高度；NFD 文件名、大小写差异及分卷 .mdd 的识别；刷新后保留结果；虚拟机桌面的各项触控手势与鼠标操作；GeoGebra 分栏比例、分隔条、收起回答与 1:1 坐标轴。上述测试同时覆盖桌面版与 iPad 版。

安装

下载 AI-Browser-iPad-unsigned.ipa，使用 AltStore、Sideloadly 等工具以自己的 Apple ID 侧载安装（iPadOS 16 或更高版本）。

---

AI Browser for iPad 3.0.1 (build 301)

This is a bug-fix and improvement update to iPad 3.0.0.

New

• Bundled Chinese document fonts: Noto Serif SC (Regular and Bold) and LXGW WenKai, used for serif and Kai text in document layout; both are distributed under the SIL Open Font License 1.1.
• Settings → About shows the iPad app version.

Fixes

• Tools glass popover in the toolbar: tapping a tool in the popover (Solve, Snap, GeoGebra, Physics, Encyclopedia, Dictionary, Code, Read Later, Scan) now opens that tool, and the "Expand to full area" button works. VoiceOver and Full Keyboard Access can also activate each tool.
• The Tools popover can be opened again right after choosing a tool, without waiting for that tool to finish loading.
• An import started while the previous Files picker is still closing is no longer ignored.
• The home page no longer keeps the CPU busy while a local model server is reachable.
• The tool icon row below the home page input is no longer shown on iPad; all tools open from the toolbar's Tools popover.
• Dictionary display area: the entry area now sizes itself to its content and is measured again after the dictionary's own scripts, fonts and images have loaded, so definitions and examples are no longer cut off. The dictionary's own style sheets, scripts and fonts are loaded as provided, and entry:// links, sound:// pronunciations and @@@LINK= redirects work; the entry background stays white. When resource files are missing, a note appears above the entry.
• Dictionary .mdd resource files: the Files picker accepts .mdx and .mdd files directly. Folder import recognises the .mdd with the same name as the dictionary as well as split files (such as .1.mdd); differences in letter case, spaces and Unicode normalisation (NFC/NFD) no longer prevent a match, and when a folder holds a single dictionary, every .mdd in it is linked to that dictionary. Dictionary settings list the .mdd files linked to each dictionary, and "Add resources…" adds .mdd files later; picking a .mdd on its own links it to the dictionary with the same name.
• Refresh: reloading a tool tab now stays on the result and runs the query again instead of returning to the tool's home page.
• Linux desktop (virtual machine) touch input now works like a trackpad: dragging one finger moves the pointer relatively without scrolling or bouncing the page; a tap is a left click, a long press is a right click, a double tap is a double click, a two-finger drag scrolls, and tap-then-drag drags with the button held. An external mouse, trackpad and Apple Pencil keep working.
• GeoGebra plots in chat: the plot takes about 55% of the width by default with 1:1 axes, and the applet is resized when the layout changes or the screen rotates. A draggable divider sits between the plot and the answer (double-click restores the default ratio), and a "Hide answer / Show answer" button collapses the answer column; the collapsed state and the ratio are remembered.

Testing

• New UI automation test: opens every tool in the popover with a real tap and confirms it opened, and tests the expand button; runs on the iPadOS 18 and iPadOS 26 simulators.
• New tests: display and automatic height of an OALD-style sample dictionary (its own style sheet, scripts, fonts, pronunciations and redirects); recognition of NFD file names, case differences and split .mdd files; reload keeping the result; every touch gesture and mouse action in the virtual machine desktop; the GeoGebra split ratio, divider, collapsible answer and 1:1 axes. These tests cover both the desktop and iPad versions.

Installation

Download AI-Browser-iPad-unsigned.ipa and sideload it with your own Apple ID using a tool such as AltStore or Sideloadly (iPadOS 16 or later).

Created by Grok

AI Browser iPad 3.0.0

AI Browser iPad 3.0.0

本版本为 iPad 版的大版本更新：引入系统 Liquid Glass 界面、会话式对话与停止生成、结构化回答、文档排版与生成助手，以及更快的模型下载；同时 Linux 镜像改为首次使用时下载，安装包更为精简。

界面

• Liquid Glass：在 iPadOS 26 上，工具栏与工具面板采用系统 Liquid Glass 材质。工具按钮以弹出面板展示全部工具，面板可展开为地址栏下方的完整工具界面；图标均以原生方式绘制。较早的 iPadOS 版本显示磨砂玻璃效果。

对话

• 会话：追问显示在上一条回答下方，对话在重启后仍完整保留；超出模型上下文时自动精简较早的轮次。
• 对话列表：侧栏列出全部对话，可搜索、重新打开、重命名与删除。
• 停止生成：回答过程中发送按钮变为停止按钮（亦可按 Esc 或 ⌘.），可同时终止回答、正在运行的工具与 Agent 步骤，已生成的内容予以保留并标注“已中止”。
• 思考过程：模型的思考内容在生成时以折叠方式实时显示。

回答格式

• 回答统一采用标题、列表、表格、引用与 LaTeX 公式。
• 化学方程式：自动排版，并检查原子守恒与电荷守恒。
• 函数图像：回答中的函数可直接在 GeoGebra 中作图。
• 单词卡片：点击即可查看词典释义。

文档与排版

• 新增排版引擎：支持 A4、A5、A3、B5、Letter、幻灯片等页面尺寸，流式与绝对定位元素、旋转、锚点、自动分页及重叠检查。
• 新增生成助手：文章与讲义、幻灯片、表格与图表、试卷与错题本、单词卡片、思维导图。
• 每份文档以卡片形式显示预览，可导出 PDF 与 HTML。在 iPad 上，PDF 由系统 PDFKit 逐页生成，每个文档页面对应一页。

Task 模式

• 助手可使用全部浏览器工具，包括标签页、工具视图、分屏、空间、稍后阅读、页面助手、截图与外观设置。每一步均以工具卡片显示，参数以简明方式呈现。

下载

• 多连接分段下载：模型文件分为多个分段并行下载，各分段均可续传，并自动检测停滞与重新分配。
• 魔搭（ModelScope）：新增为下载源，与 Hugging Face 按实测速度排序并自动切换。

GeoGebra

• 修复 GeoGebra 尚未加载完成时调用作图导致的长时间无响应；页面加载、组件加载与作图调用均设有超时，并给出明确的中文提示。

安装包与虚拟机

• Linux 虚拟机镜像（约 121 MB）不再包含于安装包中，改为首次使用虚拟机时下载，此后可离线使用。模型文件同样不随安装包提供。
• 构建流程会检查安装包内容，并确保其小于 500 MB。

安装

下载 AI-Browser-iPad-unsigned.ipa，使用 AltStore、Sideloadly 等工具以自己的 Apple ID 侧载安装（iPadOS 16 或更高版本）。

---

AI Browser iPad 3.0.0

This is a major release of the iPad version. It brings the system Liquid Glass interface, threaded conversations with a stop control, structured answers, document layout and generator assistants, and faster model downloads. The Linux image is now downloaded on first use, which keeps the app package smaller.

Interface

• Liquid Glass: on iPadOS 26 the toolbar and the tool panel use the system Liquid Glass material. The tool button opens a panel with every tool, and the panel can expand into the full tool view below the address bar. All icons are drawn natively. Earlier iPadOS versions show frosted glass.

Conversations

• Threads: follow-ups appear below the previous answer, and conversations are kept in full after a restart. Earlier turns are trimmed automatically when they exceed the model's context.
• Conversation list: the sidebar lists every conversation, which can be searched, reopened, renamed and deleted.
• Stop: while an answer is being generated, the send button becomes a Stop button (Esc or ⌘. also work). It ends the answer, any running tools and agent steps together, and keeps the partial answer marked as stopped.
• Thinking: the model's thinking is shown live in a collapsed section.

Answer format

• Answers consistently use headings, lists, tables, quotations and LaTeX formulas.
• Chemical equations are typeset and checked for conservation of atoms and charge.
• Function graphs: functions in an answer can be plotted in GeoGebra directly.
• Word cards open the dictionary entry when tapped.

Documents and layout

• New layout engine: page sizes including A4, A5, A3, B5, Letter and slides; flowing and absolutely positioned elements, rotation, anchors, automatic page breaks and overlap checks.
• New generator assistants: articles and handouts, slides, tables and charts, exam papers and mistake notebooks, word cards and mind maps.
• Each document is shown as a card with a preview and can be exported as PDF or HTML. On iPad, the PDF is produced page by page with the system PDFKit, one page per document page.

Task mode

• The assistant can use every browser tool, including tabs, tool views, split view, spaces, read later, the page assistant, screenshots and appearance settings. Each step is shown as a tool card with its arguments summarized.

Downloads

• Multi-connection downloads: model files are downloaded in parallel segments, each resumable, with automatic stall detection and reassignment.
• ModelScope is a new download source. It is ranked against Hugging Face by measured speed, with automatic fallback.

GeoGebra

• Fixed long unresponsive waits when a graph was requested before GeoGebra had finished loading. Page loading, applet loading and graph calls now have time limits with clear messages.

App package and virtual machine

• The Linux virtual machine image (about 121 MB) is no longer included in the app package. It is downloaded the first time the virtual machine is used and then works offline. Model files are not included either.
• The build checks the package contents and keeps the package under 500 MB.

Installation

Download AI-Browser-iPad-unsigned.ipa and sideload it with a tool such as AltStore or Sideloadly using your own Apple ID (iPadOS 16 or later).

Created by Grok

AI Browser iPad 2.3.1

AI Browser iPad 2.3.1

本版本修复 2.3.0 中从“文件”导入内容失败的问题：添加词典时，在“文件”选择界面点击“打开”后没有反应；其他导入入口也可能读取失败或长时间无提示。

修复

• 选择文件超过 30 秒后导入失效：添加词典的请求原先限时 30 秒。在“文件”中浏览时间较长时，请求已经超时，此后点击“打开”不再有任何反应。现在会一直等待用户完成选择。
• 统一的文件读取流程：选择文件时由系统复制所选项目（包括尚未从 iCloud 云盘下载的文件，并显示系统进度）。选择文件夹时会逐个读取其中的文件，自动下载尚未下载到本机的项目，并显示读取进度。离线百科（ZIM）、自定义系统镜像、本地模型、资料库文件、功能拓展、密码 CSV 与题目照片等所有导入入口都使用这一流程。
• 添加词典：先选择“选择文件”或“选择整个文件夹”。可一次导入包含多部词典的文件夹，每部 .mdx 词典会自动匹配同名的 .mdd 资源（包括分卷 .1.mdd、.2.mdd），以及样式、脚本、字体和图片。
• 错误提示：无法读取的文件会列出名称并说明可能的原因，不再静默失败。
• 临时文件：导入完成后删除从“文件”复制的临时副本，以节省存储空间。

测试

• 新增在 iOS 模拟器（iPadOS 18 与 iPadOS 26）上运行的自动化测试：通过真实的系统文件选择界面，逐一验证每个导入入口，并保存截图。

安装

下载 AI-Browser-iPad-unsigned.ipa，使用 AltStore、Sideloadly 等工具以自己的 Apple ID 侧载安装（iPadOS 16 或更高版本）。

---

AI Browser iPad 2.3.1

This release fixes file imports from the Files app, which failed in 2.3.0. When adding a dictionary, tapping Open in the Files picker did nothing, and other import entries could also fail to read files or go a long time without any feedback.

Fixes

• Imports stopped working after 30 seconds in the picker: the request to add a dictionary had a 30-second limit. If browsing Files took longer than that, the request had already expired, so tapping Open did nothing. The app now waits until you finish choosing.
• One file-reading flow for every import: selected files are copied by the system, including files not yet downloaded from iCloud Drive, with the system progress indicator. A selected folder is read file by file, items not yet on the device are downloaded, and progress is shown. Every import entry uses this flow: the offline encyclopedia (ZIM), custom system images, local models, knowledge-base files, feature extensions, password CSV files and problem photos.
• Adding dictionaries: choose 选择文件 (select files) or 选择整个文件夹 (select an entire folder) first. A folder holding several dictionaries can be imported at once. Each .mdx dictionary is matched with its .mdd resources (including multi-part .1.mdd and .2.mdd), as well as styles, scripts, fonts and images.
• Error reporting: files that cannot be read are listed by name with the likely cause, instead of failing silently.
• Temporary files: the copies taken from Files are deleted after import to save storage.

Testing

• New automated tests run on the iOS Simulator (iPadOS 18 and iPadOS 26). They check every import entry through the real system file picker and save screenshots.

Installation

Download AI-Browser-iPad-unsigned.ipa and sideload it with a tool such as AltStore or Sideloadly using your own Apple ID (iPadOS 16 or later).

Created by Grok

AI 浏览器 iPad 版 v2.3.0 / AI Browser for iPad v2.3.0

AI 浏览器 iPad 版 v2.3.0

本版本修复长时间使用后的连接中断、Python 与虚拟机无法启动、词典释义空白等问题，重新设计工具的打开方式，并新增拍题即解、分屏与空间、页面助手、稍后阅读、代码编辑器和自定义系统镜像。

修复

• 长时间使用后卡住并显示“bridge error”：页面与主程序之间的连接可自动恢复。断线期间的指令按编号补发，应用从后台返回后重新监听端口，缓冲区有上限；WebKit 网页内容进程被系统终止后自动重新加载页面；连接中断时显示“正在重新连接”页面，不再显示原始错误。
• Python 与 Linux 虚拟机无法启动（window.pyhost.boot、window.vmhost.beginStart 未定义）：两个隐藏的宿主页面改在页面环境中运行，并在调用前确认已就绪。
• MDX 词典（如牛津高阶）释义区空白：词条脚本在无法使用本地存储时改用内存存储，词条加载后若仍无可见文字则显示隐藏内容；导入词典时可同时选择 .mdd 资源文件，或直接选择整个词典文件夹。
• AI 释义报错 HTTP 400“invalid temperature: only 1 is allowed”：按各服务商的官方限制自动调整采样参数（例如 Kimi K2.5 及以上、OpenAI 推理模型不发送 temperature）；若服务器仍因采样参数返回 400，则去掉该参数重试一次，并记住该模型。
• GeoGebra 工具 ggb_eval 超时：首次使用时等待组件加载的时间延长至 120 秒；组件加载较慢时，稍后的调用仍可使用。
• 二维码无法扫描：改用系统相机扫描（VisionKit，旧设备使用 AVFoundation）；网址直接打开，其他内容填入地址栏。
• AI 回答中的公式显示为原始 LaTeX：数学公式与化学方程式由内置的 KaTeX（含 mhchem）离线排版。

界面

• 打开工具时直接显示工具界面，不再新建标题为“tool”的网页标签。每个工具只有一个视图，显示在标签栏左侧，可单独关闭。
• 词典查询结果在工具页内显示，占满地址栏下方区域；离线百科阅读时检索区收窄，阅读区延伸至屏幕底部。
• 每个工具主页均可“交给 AI”处理，并可切换模型；所有页面共用同一个模型选择与思考等级。思考等级按所选模型的官方参数提供，不支持调节的模型会注明。
• 工具右侧的协作侧栏可以收起。
• AI 解题的回答使用与对话相同的样式：公式排版、工具调用显示与流式输出。
• 顶部工具栏使用新的线条图标并配有中文名称。
• 界面文字统一为规范中文（“对话”“搜索”“任务”等）。

新功能

• 拍照与拍题即解：点击“拍题”即可拍摄题目，系统文字识别后自动提交解题；若所选模型支持图像，由模型转写公式。凡可添加图片之处均可直接拍照。
• 分屏与空间：两个标签页并排显示，拖动分隔线调整比例，双击交换左右；标签页可按学科（数学、物理等）归入不同空间。
• 页面助手：一键摘要网页或 PDF，并可继续提问；回答附带编号引用，点击即可滚动到原文段落并高亮。
• 稍后阅读：保存网页的完整离线副本（网页归档），无网络时亦可阅读；可与离线百科一并检索。
• 代码编辑器：带文件管理，可管理“我的代码”、Python 工作目录与虚拟机共享目录中的文件；支持语法高亮，一键运行 Python、Shell 与 HTML 文件，也可将当前文件交给 AI 讲解。
• 自定义系统镜像：在“设置 › Python 与 Linux 虚拟机”中导入 ISO、IMG 或 QCOW2 镜像并启动。镜像在 x86 模拟器（v86）中运行，磁盘按需分段读取；QCOW2 在导入时转换为原始格式。适合 FreeDOS、Alpine、TinyCore 等轻量系统。

说明

• 自定义镜像目前使用 v86 运行，仅支持 x86 系统；基于 QEMU 的解释执行方式（与 UTM 相同的思路）将在后续版本中评估。

测试

• 新增测试 41：在模拟宿主上验证工具视图、思考等级、分屏与空间、扫码、拍题即解、稍后阅读、页面助手、代码编辑器，以及在 v86 中实际启动导入的磁盘镜像。
• 单元测试新增 QCOW2（v2、v3、压缩簇）转换后逐字节一致的检查；采样参数规则与公式渲染均有单元测试。
• iPad 测试在与 iPad 相同的无 ICU、无 JIT 的 Node.js 18 上运行，Python 与 Linux 虚拟机均实际启动。

下载

文件	说明
AI-Browser-iPad-unsigned.ipa	未签名的 iPad 安装包（iPadOS 16 及以上），由 CI 构建后附加

---

AI Browser for iPad v2.3.0

This release fixes the lost connection after long use, Python and the VM failing to start, and blank dictionary entries. It redesigns how tools open, and adds photo solving, split view and spaces, a page assistant, read later, a code editor and custom system images.

Fixes

• Freezing after long use, then "bridge error": the connection between the pages and the main program now recovers on its own. Commands sent while disconnected are replayed by number, the port is listened on again when the app returns from the background, and buffers are bounded. When iPadOS terminates a WebKit content process, the page reloads. While reconnecting, a reconnect page is shown instead of the raw error.
• Python and the Linux VM failed to start (window.pyhost.boot and window.vmhost.beginStart undefined): the two hidden host pages now run in the page world, and they are confirmed ready before use.
• Blank MDX dictionary entries (such as the Oxford Advanced Learner's): entry scripts fall back to in-memory storage when local storage is unavailable, and hidden content is revealed if an entry still shows no text after it loads. When importing, .mdd resource files can be picked together with the .mdx, or the whole dictionary folder can be picked.
• AI definitions failed with HTTP 400 "invalid temperature: only 1 is allowed": sampling parameters follow each provider's official restrictions (for example, temperature is not sent to Kimi K2.5 and later or to OpenAI reasoning models). If a server still rejects a sampling parameter with a 400, the request is retried once without it and the model is remembered.
• GeoGebra tool ggb_eval timed out: on first use the tool waits up to 120 seconds for the applet to load, and later calls still work when it loads slowly.
• QR codes could not be scanned: scanning uses the system camera (VisionKit, with AVFoundation on older devices). Web addresses open directly; other content is placed in the address bar.
• Formulas in AI answers appeared as raw LaTeX: math and chemical equations are typeset offline with the bundled KaTeX (with mhchem).

Interface

• Opening a tool shows the tool directly instead of a new web tab titled "tool". Each tool has a single view, shown at the left of the tab bar, and can be closed on its own.
• Dictionary results are shown inside the tool and fill the area below the address bar. While reading the offline encyclopedia, the search area is compact and the article extends to the bottom of the screen.
• Every tool home page can hand its input to the AI ("交给 AI") and switch models. All pages share one model choice and one thinking level. Thinking levels follow the selected model's official parameters; models that cannot be adjusted say so.
• The collaboration sidebar on the right of each tool can be collapsed.
• Solver answers use the same presentation as chat: formula typesetting, tool calls and streaming.
• The top toolbar has new line icons with Chinese captions.
• Interface text is consistently in formal Chinese ("对话", "搜索", "任务" and so on).

New features

• Camera and photo solving: tap "拍题" to photograph a problem; the recognised text is submitted to the solver automatically, and a vision-capable model transcribes formulas when one is selected. Photos can be taken wherever images are accepted.
• Split view and spaces: show two tabs side by side, drag the divider to change the ratio and double-click it to swap sides. Tabs can be grouped into spaces by subject (mathematics, physics and so on).
• Page assistant: one-tap summaries of web pages and PDFs, with follow-up questions. Answers carry numbered citations; tapping one scrolls to and highlights the passage.
• Read later: save pages with complete offline copies (web archives) to read without a network, searchable together with the offline encyclopedia.
• Code editor: with a file manager for "我的代码" (my code), the Python working directory and the VM shared folder; syntax highlighting, one-tap running of Python, shell and HTML files, and handing the current file to the AI for an explanation.
• Custom system images: import and boot ISO, IMG or QCOW2 images from Settings › Python & Linux VM. Images run in the x86 emulator (v86) and the disk is read in pieces as needed; QCOW2 images are converted to raw on import. Suited to light systems such as FreeDOS, Alpine and TinyCore.

Notes

• Custom images currently run on v86 and support x86 systems only. A QEMU-based interpreter approach (as used by UTM) will be evaluated for a later release.

Tests

• New test 41 checks on the simulated host: tool views, thinking levels, split view and spaces, QR scanning, photo solving, read later, the page assistant, the code editor, and actually booting an imported disk image in v86.
• New unit check: QCOW2 images (v2, v3, compressed clusters) convert to raw byte for byte. Sampling rules and formula rendering have unit tests.
• The iPad tests run on a Node.js 18 without ICU and without a JIT, as on iPad, and Python and the Linux VM are actually started.

Download

File	Description
AI-Browser-iPad-unsigned.ipa	Unsigned iPad app (iPadOS 16 or later), attached after the CI build

Created by Grok

AI Browser for iPad v2.2.2

AI 浏览器 iPad 版 v2.2.2

修复版本：解决 v2.2.1 在真机上启动时报“主程序（main.js）加载失败：The "big5" encoding is not supported”的问题。

修复：主程序加载失败（big5 编码不受支持）

• 原因：内置的 Node.js 运行时（nodejs-mobile）在 iOS 上不含 ICU 国际化数据，其文本解码器（TextDecoder）只支持 UTF-8 和 UTF-16LE。词典模块在加载时即创建 BIG5 与 GB18030 解码器，因此抛出异常，导致整个主程序无法加载。模拟宿主此前使用带完整 ICU 的 Node.js，未能暴露该问题。
• 修复：在主程序之前加载文本解码兼容层。运行时自带解码器不支持的编码（BIG5、GB18030、GBK、GB2312、Shift_JIS、EUC-JP、EUC-KR、UTF-16BE、Windows-125x、ISO-8859 系列、KOI8 等）以及 fatal 选项，改由纯 JavaScript 实现的 iconv-lite（MIT 许可）解码，支持流式解码。
• 词典解析模块改为首次使用时才加载；即使其加载失败，也只影响词典功能，不再影响主程序启动。
• 以 BIG5、GBK 等编码返回的网页，其正文提取同样恢复正常。

兼容性审查

已在与 iPad 相同条件（Node.js 18、无 ICU、无 JIT）下检查主进程代码及其依赖：不使用 Intl；排序、数字与日期格式化、Unicode 规范化在无 ICU 时不会出错；除文本解码器外未发现其他启动时故障。

测试与持续集成

• 持续集成新增以 --with-intl=none 方式编译的 Node.js 18（与 nodejs-mobile 相同），模拟宿主上的 iPad 测试改为在该运行时上、以无 JIT 方式运行。
• 新增设备运行时检查：在无 ICU、无 JIT 条件下验证各类编码的解码结果，并逐一加载主进程依赖的模块。
• 本地已在上述条件下通过全部 23 项 iPad 模拟宿主测试与词典测试。
• 构建流程在不同 macOS 构建机镜像上自动选择可用的 Xcode。

下载

文件	说明
AI-Browser-iPad-unsigned.ipa	未签名的 iPad 安装包（iPadOS 16 及以上），由 CI 构建后附加

---

AI Browser for iPad v2.2.2

A fix release. It resolves the startup failure of v2.2.1 on real devices: "主程序（main.js）加载失败: The "big5" encoding is not supported".

Fix: main program failed to load (big5 encoding not supported)

• Cause: the embedded Node.js runtime (nodejs-mobile) is built without ICU internationalisation data on iOS, so its TextDecoder supports only UTF-8 and UTF-16LE. The dictionary module creates BIG5 and GB18030 decoders when it loads, which threw an exception and prevented the whole main program from loading. The simulated host previously used a Node.js with full ICU and did not expose the problem.
• Fix: a text decoding compatibility layer is loaded before the main program. Encodings the built-in decoder does not support (BIG5, GB18030, GBK, GB2312, Shift_JIS, EUC-JP, EUC-KR, UTF-16BE, Windows-125x, the ISO-8859 family, KOI8 and others), as well as the fatal option, are decoded by iconv-lite (MIT licence), written in plain JavaScript, including streamed decoding.
• The dictionary parser is now loaded on first use. Should it fail to load, only the dictionary feature is affected; the main program still starts.
• Text extraction from web pages served in BIG5, GBK and similar encodings works again as well.

Compatibility review

The main-process code and its dependencies were checked under the same conditions as on iPad (Node.js 18, no ICU, no JIT): Intl is not used; sorting, number and date formatting and Unicode normalisation do not fail without ICU; apart from the text decoder, no other startup failure was found.

Tests and continuous integration

• CI now builds Node.js 18 with --with-intl=none (as nodejs-mobile is built), and the iPad test on the simulated host runs on that runtime without a JIT.
• A new device runtime check verifies decoding results for many encodings and loads every module the main process depends on, without ICU and without a JIT.
• All 23 iPad tests on the simulated host and the dictionary tests passed locally under these conditions.
• The build selects an available Xcode on every macOS runner image.

Download

File	Description
AI-Browser-iPad-unsigned.ipa	Unsigned iPad app (iPadOS 16 or later), attached by CI after the build

Created by Grok

AI Browser v5.7.0

AI 浏览器 v5.7.0

图形桌面镜像与电脑操作工具

Linux 虚拟机新增可选的图形镜像：Debian 13（arm64）+ XFCE 桌面，内置 Firefox ESR、终端、文件管理器、文本编辑器，以及 Python 科学计算库、git、ripgrep 和编译工具。在 设置 → Python 与 Linux → Linux 虚拟机 → 系统镜像 中选择“图形桌面（Debian 13 + XFCE）”即可。镜像在首次使用时下载（约 990 MB），之后保存在本机。命令行镜像（Alpine）的磁盘会保留，两个镜像可随时切换，磁盘互相独立。

AI 新增 7 个电脑操作工具，可以像人一样操作虚拟机里的图形程序：

工具	作用
vm_desktop	启动、查看状态、关闭图形会话，或调整屏幕分辨率
vm_screenshot	截取虚拟机桌面
vm_click	在指定坐标单击、双击或右键
vm_type	输入文字（支持中文等 Unicode 字符）
vm_key	按键与组合键，如 Return、ctrl+l、alt+F4
vm_scroll	在指定位置滚动
vm_drag	拖动

每次操作后都会返回新的截图。截图会发送给支持图像输入的模型：OpenAI 兼容接口以图片消息发送，Anthropic 接口放在工具结果中。只保留最近两张截图，以控制上下文长度。不支持图像的模型只会收到文字结果，并被告知看不到截图。

Linux 桌面查看器

新增“Linux 桌面”页面（设置中的“打开桌面”按钮）：实时显示虚拟机画面，你可以直接点击、滚动和用键盘输入，与 AI 一起操作同一个桌面。

实现

• 图形镜像与命令行镜像使用同一个内核和 initramfs，由 systemd 启动同一个虚拟机代理；Codex 风格的 9 个工具（v5.6.0）在图形镜像中同样可用，vm_pkg 使用 apt，vm_service 使用 systemd。
• 桌面运行在 Xvfb 上（无需虚拟显卡），截图直接读取帧缓冲并在虚拟机内转换为 PNG，输入通过 xdotool 完成。
• 镜像在 Linux 上通过 QEMU 启动 Alpine 虚拟机、在其中运行 debootstrap 构建（vm/debian/build-debian-gui.js），不需要 binfmt。

测试

• 新增测试 38：在 QEMU 上启动 Debian 图形镜像，截图、点击（验证鼠标位置）、在终端中输入命令并回车（验证命令真的执行）、滚动、拖动、组合键，拒绝无效按键，并验证切回命令行镜像后两块磁盘互不影响；截图送给视觉模型的消息格式也有单元测试。
• 测试 38 实测（Linux 主机上的 QEMU，TCG 模拟 arm64，比 Mac 上的 vfkit 慢得多）：启动 29.4 秒，启动桌面 1.1 秒，桌面绘制完成 16.5 秒，截图 0.9 秒，点击 17.5 秒（含等待终端窗口出现），输入 2.3 秒，按键 2.0 秒，滚动 1.5 秒。所有图形测试通过。

下载

文件	说明
AI-Browser-mac-arm64.zip	未签名的 AI Browser.app（macOS arm64）
install.command / install.sh	解压 → 复制到“应用程序” → xattr -cr → 启动
AI-Browser-source.zip	源代码
README.md / README-zh.md	英文 / 中文说明
linux-vm-arm64-debian-xfce.img.gz	图形镜像（首次选择时由应用自动下载）
linux-vm-x86-gui.tar	iPad 版轻量图形镜像（首次选择时由 iPad 应用自动下载）

---

AI Browser v5.7.0

Graphical desktop image and computer-use tools

The Linux VM has a new optional graphical image: Debian 13 (arm64) with the XFCE desktop, Firefox ESR, a terminal, a file manager and a text editor, plus the Python scientific stack, git, ripgrep and build tools. To use it, open Settings → Python & Linux → Linux VM → System image and choose "Graphical desktop (Debian 13 + XFCE)". The image is downloaded on first use (about 990 MB) and then kept on your Mac. The disk of the command-line image (Alpine) is kept too, so you can switch back and forth at any time; each image has its own disk.

The AI gets seven computer-use tools for working with graphical programs in the VM:

Tool	What it does
vm_desktop	Start the graphical session, check its status, stop it, or change the screen size
vm_screenshot	Take a screenshot of the VM desktop
vm_click	Click, double-click or right-click at a position
vm_type	Type text (Unicode, including Chinese)
vm_key	Press keys and shortcuts such as Return, ctrl+l, alt+F4
vm_scroll	Scroll at a position
vm_drag	Drag

Each action returns a fresh screenshot. Screenshots go to models that accept images: as an image message for OpenAI-compatible APIs, and inside the tool result for Anthropic. Only the two newest screenshots are kept, to limit the context size. Models without image input get the text results only and are told that they cannot see the screenshots.

Linux desktop viewer

A new "Linux desktop" page (the "Open desktop" button in Settings) shows the VM screen live. You can click, scroll and type in it, and work on the same desktop as the AI.

How it works

• The graphical image uses the same kernel and initramfs as the command-line image, and systemd starts the same VM agent. The nine Codex-style tools from v5.6.0 work in it too; vm_pkg uses apt and vm_service uses systemd.
• The desktop runs on Xvfb, so no virtual GPU is needed. Screenshots read the framebuffer and are converted to PNG inside the VM; input goes through xdotool.
• The image is built on Linux by booting the Alpine VM in QEMU and running debootstrap inside it (vm/debian/build-debian-gui.js), so binfmt is not needed.

Tests

• New test 38 boots the Debian graphical image on QEMU. It takes screenshots, clicks (and checks the pointer position), types a command into a terminal and presses Return (and checks that the command really ran), scrolls, drags, sends shortcuts, rejects invalid keys, and checks that the two image disks stay separate after switching back. The image message formats for vision models also have unit tests.
• Measured in test 38 (QEMU on a Linux host, emulating arm64 with TCG, which is much slower than vfkit on a Mac): boot 29.4 s, desktop start 1.1 s, desktop drawn 16.5 s, screenshot 0.9 s, click 17.5 s (including waiting for the terminal window), type 2.3 s, key 2.0 s, scroll 1.5 s. All graphical tests pass.

Download

File	Description
AI-Browser-mac-arm64.zip	Unsigned AI Browser.app (macOS arm64)
install.command / install.sh	Unzip → copy to Applications → xattr -cr → launch
AI-Browser-source.zip	Source code
README.md / README-zh.md	English / Chinese readme
linux-vm-arm64-debian-xfce.img.gz	The graphical image (downloaded by the app when you first choose it)
linux-vm-x86-gui.tar	The light graphical image for iPad (downloaded by the iPad app when you first choose it)

Created by Grok