TabCtrl

Tracking since Apr 30, 2026.

Changelog

• May 7, 2026
  description

  TabCtrl 是一款面向高级用户和团队内测场景的浏览器 AI 控制扩展。它允许用户接入自己配置的模型服务和 API Key，让 AI Agent 在用户当前浏览器标签页中观察页面、理解任务、生成计划，并在用户授权范围内执行点击、输入、滚动、标签页管理、页面读取和结果验证等操作。
  用户安装 TabCtrl 的主要价值是提升复杂网页任务的执行效率。相比普通聊天机器人，TabCtrl 可以直接在真实网页环境中工作，读取当前页面结构，识别表单、按钮、链接和跨 iframe 内容，并按照用户目标逐步完成任务。它适用于网页资料整理、后台操作辅助、表单填写、内容编辑、跨页面信息处理、文档系统操作、以及需要模型视觉或推理能力参与的复杂浏览器任务。
  TabCtrl 采用 Bring Your Own Key 模式。用户需要自行配置模型端点和 API Key，扩展不会内置公共模型账号，也不会把用户 API Key 上传到开发者服务器。模型配置、站点策略、技能、教学案例和任务状态保存在用户本机 Chrome 配置中。
  为了完成浏览器自动化任务，TabCtrl 需要读取和操作网页内容。`<all_urls>` 权限用于让用户在任意自己打开并授权使用的网页上调用 TabCtrl，而不是限制在单个固定网站。`activeTab`、`tabs`、`tabGroups`、`sidePanel` 和 `scripting` 用于管理当前任务标签页、打开侧边栏、注入内容脚本并执行浏览器操作。`storage` 和 `unlimitedStorage` 用于保存模型配置、任务上下文、技能、教学案例和用户策略。`notifications` 与 `offscreen` 用于任务完成提醒和提示音。`debugger` 权限用于在复杂网页编辑器、富文本输入框或常规 DOM 输入无法可靠工作的场景中提供更稳定的文本输入能力。`nativeMessaging` 用于实验室功能中的本机桥接，默认关闭，只有在用户显式开启并批准操作时，才会调用本机 allowlist 中的 CLI 工具。
  TabCtrl 的设计目标是让用户始终保持控制权。敏感操作会请求用户批准；计划模式允许用户先审查 AI 的执行计划；本机桥接默认关闭并需要审批；用户可以随时停止任务。扩展不会绕过网站登录状态，而是在用户已登录的真实浏览器环境中，以用户明确给出的任务目标为依据工作。
  用户应该安装 TabCtrl，如果他们希望把 AI 从“只能给建议的聊天窗口”扩展为“可以在真实网页中辅助完成任务的浏览器控制面板”，并且愿意自行配置可信模型服务、理解自动化权限，并在关键操作前保持人工确认。
  English Version
  TabCtrl is an AI browser-control extension designed for advanced users and team testing scenarios. It lets users connect their own model endpoints and API keys, then allows an AI agent to observe the current browser tab, understand the user’s task, draft a plan, and perform approved browser actions such as clicking, typing, scrolling, reading page content, managing tabs, and verifying results.
  The main value of TabCtrl is improving the efficiency of complex web-based tasks. Unlike a regular chatbot, TabCtrl can work inside the user’s real browser environment. It can inspect page structure, identify forms, buttons, links, and iframe content, and execute step-by-step actions toward the user’s goal. It is useful for web research, admin workflows, form filling, content editing, multi-page information processing, document-system operations, and tasks that benefit from model reasoning or visual understanding.
  TabCtrl follows a bring-your-own-key model. Users configure their own model provider, endpoint, and API key. The extension does not include a shared model account, and it does not upload the user’s API key to the developer’s server. Model profiles, site policies, skills, teaching cases, and task state are stored locally in the user’s Chrome profile.
  To perform browser automation, TabCtrl needs permission to read and interact with web pages. The `<all_urls>` permission allows users to use TabCtrl on any page they choose to open and authorize, instead of limiting the extension to one specific website. `activeTab`, `tabs`, `tabGroups`, `sidePanel`, and `scripting` are used to manage task tabs, open the side panel, inject content scripts, and perform browser actions. `storage` and `unlimitedStorage` are used to save model profiles, task context, skills, teaching cases, and user policies. `notifications` and `offscreen` are used for task completion notifications and a short completion sound. The `debugger` permission is used to provide more reliable text input in complex editors, rich-text fields, or pages where standard DOM typing is not sufficient. `nativeMessaging` powers the optional Lab native bridge. It is disabled by default and can only invoke allowlisted local CLI tools after the user explicitly enables the feature and approves the action.
  TabCtrl is designed to keep the user in control. Sensitive actions require user approval. Plan mode lets the user review the agent’s intended steps before execution. Native bridge access is disabled by default and requires approval. Users can stop a task at any time. The extension does not bypass website authentication; it works within the user’s existing logged-in browser session and follows the task goal explicitly provided by the user.
  Users should install TabCtrl if they want to extend AI from a chat-only assistant into a browser control surface that can help complete real web tasks, while still keeping model choice, API keys, permissions, and critical approvals under their own control.

  TabCtrl 让你用自己的模型 API Key，把浏览器变成可执行的工作台。它能读取页面、点击、输入、跨 frame 操作、管理标签页，并按你的指令完成多步任务。
  核心特性
  自带模型：兼容多种协议。
  多模型接力：可同时配置 primary、vision、reasoning、fallback 角色，按任务自动路由——视觉任务交给 vision 模型，复杂规划交给 reasoning 模型。
  浏览器工具：read_page、click、type、scroll、find、navigate、screenshot、标签页操作，覆盖日常网页自动化。
  Plan 模式：复杂任务先生成计划并请你审批，按步执行，偏离前主动确认。
  教学案例录制：示范一次操作，模型在后续相似任务中参考你的处理方式。
  Skill 系统：导入文件夹形式的 SKILL.md，按需加载专门任务说明。
  任务标签组隔离：模型只能操作当前任务标签组内的页面，不会乱动你的其它工作。
  中英文界面自适应。
  隐私
  API Key 与对话仅保存在你的浏览器本地存储，不经过 TabCtrl 服务器。
  详细政策见隐私页面。
  适合谁
  希望用自己的模型与端点完成网页任务的开发者与高级用户。
  需要在飞书、文档、表格、研究等场景中复用稳定操作流程的团队。
  TabCtrl turns your browser into an executable workspace driven by your own model API keys. It reads pages, clicks, types, works across frames, manages tabs, and carries out multi-step tasks on your behalf.
  Highlights
  Bring your own keys: Compatible with multiple protocols
  Multi-model relay: configure primary, vision, reasoning, and fallback roles. Visual results route to a vision model; hard planning routes to a reasoning model.
  Browser tools: read_page, click, type, scroll, find, navigate, screenshot, and tab operations cover everyday web automation.
  Plan mode: for complex tasks, the agent drafts a short plan, waits for your approval, and asks again before deviating.
  Teaching cases: demonstrate an action once and let the model reference your approach in similar tasks later.
  Skills: import folders containing a SKILL.md to load task-specific instructions on demand.
  Task tab-group isolation: the agent can only see and act on tabs inside the current TabCtrl task group, leaving the rest of your browser alone.
  Bilingual UI (English / 中文).
  Privacy
  API keys and conversations are stored locally in your browser. They do not pass through any TabCtrl server.
  See the privacy policy for details.
  Who it's for
  Developers and power users who want to drive web tasks with their own model and endpoint.
  Teams that need repeatable workflows across docs, sheets, Feishu/Lark, and research pages.

• May 7, 2026
  category

  productivity/workflow

  productivity/tools

Permissions & access

Permissions

activeTabscriptingtabstabGroupsstoragesidePanelwebNavigationunlimitedStoragealarmsnotificationsoffscreendebuggernativeMessaging

Host access

<all_urls>

Screenshots