GitHub上，一个叫OpenWorker的项目上线不到48小时就拿到了3700个Star。项目作者是吴恩达——那个在Coursera上教过几百万人学机器学习的吴恩达。
7月24日，吴恩达在LinkedIn正式宣布开源OpenWorker——一个定位为“个人桌面AI同事”的智能体。它不是又一个聊天机器人，而是一个能在你电脑上跨应用自主完成任务的Agent。润色文件、准备客户简报、梳理日程、起草报告、处理Slack警报——这些以前需要你手动切换三五个应用完成的事，现在一句话交代给它就行。
三个关键设计：为什么OpenWorker不一样
桌面Agent这个概念并不新鲜，微软的Copilot+、苹果的Apple Intelligence都有类似的定位。但OpenWorker有三个设计值得关注，它们代表了Agent发展的一个新方向。
第一个是100%开源，MIT协议。这意味着你可以自由地改代码、做二次开发、甚至商用。微软和苹果的桌面Agent都是闭源的黑盒子，你不知道它在干什么、数据去哪了。OpenWorker把所有代码都摊开在你面前——Agent循环、工具调用、对话记录，全透明。
第二个是本地优先。Agent的运行循环、对话记录、API密钥、工具token——这些敏感数据默认都存在你本地电脑上，不需要上传到云端。吴恩达团队在项目首页明确写了“本地优先”的设计原则，这在当前Agent产品几乎全是云端SaaS的大环境下显得格外突出。隐私焦虑越来越重的今天，“数据不出本地”可能会成为桌面Agent的核心竞争力。
第三个是不绑定模型。OpenWorker基于aisuite库构建，这是一个与提供商无关的Python框架，通过统一API支持多种大模型。你可以用GPT 5.6 Sol、Claude Fable、Gemini 3.6，也可以用Kimi、GLM、DeepSeek这些国产模型，甚至可以通过Ollama跑本地开源模型。用户自带API Key，OpenWorker不赚模型差价。
Agent入口之争：桌面是下一个战场
回顾Agent的入口演进，你会看到一条清晰的移动轨迹。
第一站是聊天框。ChatGPT、Claude、文心一言——所有Agent的起点都是对话框。你在框里输入指令，AI输出结果。简单直接，但能做的事情有限，而且需要用户主动打开应用。
第二站是IDE。Cursor、GitHub Copilot、Claude Code把Agent放进了代码编辑器里。程序员是第一批真正把Agent用起来的人，因为编程场景有明确的目标、清晰的反馈、大量的工具可以调用。Agent在IDE里证明了自己的价值。
第三站正在到来——桌面。OpenWorker代表的是Agent从垂直场景（聊天、编程）走向通用场景的尝试。你的文件系统、日历、邮件、办公软件、项目管理工具——所有这些日常工作的入口都在桌面上。如果Agent能在桌面上把这些工具串起来，它就不只是一个编程助手或聊天机器人，而是真正的“工作伙伴”。
当然，桌面Agent的挑战也很大。操作系统权限、应用之间的互操作、数据安全、指令理解的准确性——每一个都是硬骨头。OpenWorker目前只支持macOS，Windows版还在路上，Linux更是遥遥无期。但它的出现标志着：Agent的入口之争已经从浏览器和IDE，扩展到了整个操作系统桌面。
谁能占领桌面，谁就可能成为AI时代的“操作系统级入口”。这场战争才刚刚开始。
明天见。
On GitHub, a project called OpenWorker picked up 3,700 stars in less than 48 hours. Its author is Andrew Ng — the same Andrew Ng who taught machine learning to millions on Coursera.
On July 24, Ng officially announced the open-source launch of OpenWorker on LinkedIn — an agent positioned as a “personal desktop AI coworker.” It's not another chatbot. It's an agent that can autonomously complete tasks across applications on your computer. Polishing documents, preparing client briefs, organizing schedules, drafting reports, handling Slack alerts — things that used to require manually switching between three to five apps can now be handed off with a single sentence.
Three Key Design Choices: Why OpenWorker Is Different
The desktop agent concept isn't new. Microsoft's Copilot+ and Apple's Apple Intelligence have similar positioning. But OpenWorker has three design choices worth paying attention to — they represent a new direction for agent development.
First: 100% open source, MIT license. This means you can freely modify the code, build on it, even use it commercially. Microsoft's and Apple's desktop agents are closed-source black boxes — you don't know what they're doing or where your data goes. OpenWorker lays all its code bare: the agent loop, tool calls, conversation history — everything is transparent.
Second: local-first. The agent's runtime loop, conversation history, API keys, tool tokens — all this sensitive data stays on your local computer by default, no cloud upload required. The Ng team explicitly states “local-first” as a design principle on the project homepage. This stands out in today's environment where almost all agent products are cloud SaaS offerings. In an era of growing privacy anxiety, “data stays local” could become the core competitive advantage of desktop agents.
Third: no model lock-in. OpenWorker is built on the aisuite library, a provider-agnostic Python framework that supports multiple LLMs through a unified API. You can use GPT 5.6 Sol, Claude Fable, Gemini 3.6, or Chinese models like Kimi, GLM, DeepSeek — you can even run local open-source models via Ollama. Users bring their own API keys; OpenWorker doesn't take a cut on model fees.
The Agent Entry Point War: Desktop Is the Next Battlefield
Looking back at the evolution of agent entry points, you can see a clear trajectory.
Stop one was the chat box. ChatGPT, Claude, Wenxin Yiyan — every agent starts in a dialog box. You type instructions, the AI outputs results. Simple and direct, but limited in what it can do, and it requires the user to actively open the app.
Stop two was the IDE. Cursor, GitHub Copilot, Claude Code put agents inside code editors. Programmers were the first group to really put agents to work, because programming has clear goals, crisp feedback, and lots of tools to call. Agents proved their value in the IDE.
Stop three is arriving now — the desktop. OpenWorker represents an attempt to move agents from vertical scenarios (chatting, coding) toward general scenarios. Your file system, calendar, email, office software, project management tools — all these daily work entry points live on the desktop. If an agent can string these tools together from the desktop, it stops being just a coding assistant or chatbot and becomes a real “work partner.”
Of course, desktop agents face huge challenges too. OS permissions, inter-app interoperability, data security, instruction understanding accuracy — each one is a hard problem. OpenWorker currently only supports macOS; the Windows version is still on the way, and Linux is nowhere in sight. But its appearance marks something: the battle for the agent entry point has expanded from browsers and IDEs to the entire OS desktop.
Whoever captures the desktop could become the “OS-level entry point” of the AI era. This war has only just begun.
See you tomorrow.
当Agent从聊天框和IDE走向桌面，它就不再是一个工具插件，而是在争夺AI时代的操作系统级入口。
—— Dawn Vision编辑部
When agents move from chat boxes and IDEs to the desktop, they stop being tool plugins and start competing for the OS-level entry point of the AI era.
—— The Dawn Vision Editorial Desk
Andrew Ng · OpenWorker · desktop agent · open source · local-first · multi-model · tool integration · AI entry point
Sources · 信源 Sources
本文基于 Dawn Vision 认知引擎处理的 8 个源信号生成，经编辑部人工审核。素材来源：IT之家、量子位、GitHub。
This article was generated by the Dawn Vision cognitive engine processing 8 source signals, with human editorial review. Sources: IT Home, QbitAI, GitHub.
