OpenWorker

An open-source, local-first desktop AI coworker that delivers finished work, not conversations

Agent Universal Agent

https://github.com/andrewyng/openworker

Visit Website

In-depth Report

In July 2026, AI education leader Andrew Ng and former Amazon AGI head Rohit Prasad jointly open sourced the desktop AI Agent "OpenWorker" under the MIT license. Unlike traditional "chat-and-answer" AI, OpenWorker's core positioning is to "deliver finished products rather than conversations" - users only need to describe the desired results (such as "prepare client briefings"), and it can break down the tasks by itself, execute them across local files and connected applications, and finally hand over a file that can be opened and shared directly. After the project was launched on GitHub, it quickly received more than 3,500 stars, marking the emergence of a third position in the battle for desktop Agent entrance: open source, local first, and model neutral.
Andrew Ng is the founder of DeepLearning.AI, co-founder of Coursera, and adjunct professor of the Department of Computer Science at Stanford University. He once led the Google Brain project and served as Baidu's chief scientist. According to his personal official website, more than 7 million people have learned artificial intelligence through his courses. The open source of OpenWorker this time continues his consistent philosophy of "lowering the threshold for using AI" - only this time the delivery is not a course, but an AI colleague. Co-developer Rohit Prasad is equally hefty. He joined Amazon in 2013 and was the early chief scientist of Alexa. He has led the Amazon AGI team and led the Nova basic model series since 2023. He resigned at the end of December 2025 during the restructuring of Amazon's AI architecture. About seven months after leaving Amazon, he appeared as a co-developer on the open source project, his first public move since leaving a major company. OpenWorker was initially developed in the aisuite warehouse, a multi-model interface library that was previously open sourced by Andrew Ng's team, and then became independent and established a separate warehouse. The project entered the public beta phase on July 23, 2026. The current macOS (Apple Silicon) version is signed and supports automatic updates. The Windows 10/11 version is available but not yet code signed.
The core difference between OpenWorker and ordinary AI chatbots is the "delivery closed loop". The traditional process of using large models is to let AI analyze requirements and generate text, and then users find files, check data, adjust formats, and finally send the results in Slack, mailbox, or calendar. The goal of OpenWorker is to open up this link - users only need to describe the desired results, and it will disassemble the task itself, call files, terminals and connected applications on the computer, and finally deliver a finished product that can be used directly. In terms of connectivity capabilities, OpenWorker has more than 25 built-in connectors for office tools, covering mainstream tools such as GitHub, Slack, Jira, Notion, Linear, HubSpot, Outlook, Gmail, and Google Calendar. It also supports access to any compatible tool through the MCP protocol. Users can trigger @OpenWorker tasks in the Slack channel, and it will reply the results to the Slack thread after the machine completes the work. Scheduled automation is another important feature. OpenWorker supports periodic tasks such as morning report, weekly report, channel watch, etc., and leaves a complete running record for each run. This can save a lot of repetitive operations for knowledge workers who need routine information aggregation. At the architectural level, OpenWorker uses Tauri 2 + React desktop shell, paired with a local Python Agent Server. The Agent engine is based on Ng's previously open source aisuite library. According to GitHub warehouse data, the project contains approximately 119 Python back-end files, 149 TypeScript/TSX front-end files, and 78 back-end test modules. The project integrity is outstanding among similar open source projects.At the model level, OpenWorker is not bound to any vendor. Users can access cutting-edge models such as OpenAI (GPT-5.6 Sol/Terra/Luna), Anthropic (Claude Fable/Mythos/Opus/Sonnet), Google Gemini 3.6, DeepSeek, Qwen, Mistral, Grok, etc. by bringing their own API keys. They can also run completely local models through Ollama. The curated model list covers approximately 30 tool call-validated models, and even supports switching between different models in the same task to balance cost and quality.
OpenWorker is completely free under the MIT open source license and has no paid tiers. Users only bear the cost of the model API of their choice - running with native Ollama models costs zero. For open source projects, this pricing strategy lowers the barrier to try, but it also means there are no managed cloud options, no enterprise SLA, and no compliance certification, suitable for users and teams who have the ability to maintain themselves.
Judging from multiple feedbacks, the overall evaluation of OpenWorker in the industry is relatively positive. The core highlights focus on several aspects: first, the product definition of "delivering finished products" meets the real needs of knowledge workers for AI to "live" rather than "chat"; second, the trinity combination of open source + local priority + model neutrality is unique among existing desktop Agent products; third, the engineering implementation of the typed permission engine is recognized by the developer community. At the same time, early reviews also pointed out a number of limitations. The Windows version is not yet code signed and will trigger a SmartScreen warning when installed. The project is currently in the beta stage, and the stability of the connector ecosystem and large-scale user verification are still unknown. In addition, users need to bring their own API keys or deploy Ollama by themselves, and the barrier to entry cannot be ignored for non-technical users.
The interpretation of OpenWorker by many technology media focuses on the same judgment: it is stuck in a previously empty position. In the first half of 2026, the battle for the entry of desktop agents will unfold along two routes - Anthropic's Claude Cowork works as a desktop clerk within the security boundary and deeply binds its own model; the open source OpenClaw takes the deep system access route, is self-hosted, has a flexible model but is driven by the community. OpenWorker's open source, local first, model neutrality, and product definition of "delivering finished products" rather than "assisting conversations" just form the third route. The significance of this position lies in the neutrality of the model manufacturer. Cowork products are essentially distribution channels for model companies, while OpenWorker turns models into replaceable components. Its support for domestic models—Kimi, GLM, and DeepSeek go directly into the featured list—means that domestic developers can use it as a daily Agent base, not just a client for overseas models. However, some analysts pointed out that Andrew Ng's projects have always had educational and ecological attributes: aisuite is an interface standard, and OpenWorker is more like its reference implementation, giving developers a "same look" and is still far from a mass product. The degree of polishing of open beta and the stability of the connector ecosystem will take time to verify.
At the security level, a local Agent with terminal access permissions and file write permissions is itself a persistent attack surface. Although OpenWorker's permission engine has a typed design - divided into four risk levels of read-only, local writing, command execution, and external side effects, and supports five permission modes - the approval pop-up window before performing important operations is essentially still the last line of defense to prevent malicious instructions. Hint injections hidden in crawled web pages, compromised connector tokens, and malicious MCP servers can all bypass this mechanism. Another risk lies in the beta status of the project. The Windows version is not yet code-signed, and making the cross-platform experience consistent is still unfinished work. Whether the 25 built-in connectors can cover the real workflow of the enterprise and the actual experience of MCP access will take longer to verify. In addition, the increasing functional overlap with OpenClaw is also a competitive variable worth watching. Both are open source, both are model-neutral, and both emphasize local operation. The differences will converge into interaction forms and engineering governance. It remains to be seen whether OpenWorker can leverage Ng’s community influence to build a contributor ecosystem.
OpenWorker is most suitable for the following types of users: knowledge workers who value data privacy (run locally, data does not leave the device); developers who are willing to toss (open source, auditable, and secondary development); users who want to flexibly switch between multiple models (with their own API keys, models can be replaced); teams that need to automate office tasks on a regular basis. Scenarios that are less suitable include: non-technical users who need zero setup out of the box (API key or Ollama needs to be configured); organizations that require enterprise-level compliance and security certification; enterprises that require multi-user team collaboration and centralized management capabilities; users who mainly use Windows or Linux and do not want to bother (macOS is currently the main platform).
The value of the release of OpenWorker is not only a new Agent product, but also a clear statement by Andrew Ng on the development direction of desktop Agents: AI Agents cannot stay in the chat box forever, but must truly participate in work, deliver results, and return the right to model selection, tool calling, and final decision-making to the user. For teams that are evaluating Agent tools, OpenWorker's approval permissions model can serve as a baseline against which to measure Agent products on the cloud. Its degree of polishing in the beta phase and the progress of community ecology construction will determine whether it can evolve from a "reference implementation" to a "mass product."

User Reviews

Lawrence_Wilson_99

—

Windows 版本装了，但 SmartScreen 一直弹警告，搞了半天才装上。虽然理解 beta 阶段还没签名，但体验确实打了折扣。等正式版吧。

CryptoTr_ader885

—

吴恩达果然是吴恩达，OpenWorker 这个「交付成品」的思路太对了。之前每次用 AI 都要在各软件之间来回倒腾，现在一句话就搞定，省了中间环节。

Gabriel.Bailey_7

—

太香了，已经回不去了。

Heather.Price_775

—

看了一圈代码，权限引擎的设计是最大的亮点。把工具调用分成 read / write_local / exec / external 四类，再配五种权限模式，比 OpenClaw 那种简单的全部允许或全部拒绝合理得多。

Thomas_Miller_88

—

刚在 Mac 上装好 OpenWorker，配置了 Claude API Key，试了一下让它帮我整理下周的日程冲突。说实话第一次跑的时候有点担心它会不会乱改日历，结果它在执行前弹了个确认框，列了具体的改动清单让我审一遍。比起 Cowork 那种黑盒操作，这种感觉踏实不少。

JMitchellIII

—

期待 Windows 版本完善！

realLyubomudrSviridovskiy_88

—

yyds！吴恩达出品必属精品。

MsFrankieDunne_88

—

之前用 OpenClaw 总担心安全问题，OpenWorker 的权限引擎给了我没尝试过的安心感。把每个工具调用都做了风险评级再决定是否放行，而不是简单粗暴的全放权或全禁止，这种精细化的管控才是一个桌面 Agent 该有的样子，尤其是当它拥有你终端的访问权限时，每一步都慎重才是对的。

JoshuaPerez007

—

GitHub 上已经有汉化包了，叫 openworker-zh-CN，一位开发者做的。覆盖了大概 80% 的界面，侧边栏、设置页、连接器都汉化了。对英文不好的用户来说，这个贡献真的很关键。

杜博雅

—

有没有人试过在 Ollama 上跑本地模型？我用 Qwen 2.5 试了一下，执行速度和响应质量都还不错，关键是数据完全不出本机，这比用云端 API 放心多了。

i1gj2jx

—

本地运行这点就秒杀 Cowork 了。

Kimberly_Morgan369

—

说实话上手门槛还是有点高的。先要装客户端，然后是 Python 环境，还要自己搞 API Key。对技术小白来说，光是配置这一步就劝退了。但如果你是开发者，那它就是神器。

CherylPowell_2021

—

免费的，还要什么自行车。

WElin_dev

—

试了下让它从 HubSpot 拉客户数据然后生成简报，最后产出的是一个可以直接用的 HTML 文件。之前用 GPT 做同样的事情，我需要手动跑 CRM 导出、再复制粘贴到提示词里，OpenWorker 省了中间至少三步。

安然297

—

最让我惊喜的是 Slack 集成。在频道里 @OpenWorker 就能触发任务，它在桌面端跑完活把结果回复到线程里，这个体验很自然，不用在聊天和桌面应用之间切来切去。

Kevin.ThompsonQ

—

仔细读了一下 secrets.py 的代码，发现连接器的访问令牌目前还是以 0600 权限的 JSON 文件存着的，官方说后续会切到系统密钥链。正式用之前建议注意一下这个点。

FlamingoFiRich_ter

—

3.7K stars 几天就到了，吴恩达的影响力确实不是盖的。不过说真的，这项目目前还处在「参考实现」的阶段，连接器的稳定性和 bug 修复速度才是决定它能走多远的关键。

EliWalker

—

在 Slack 里 @OpenWorker 这个功能太香了。我在运营频道设了一个每天早上 9 点自动生成数据简报的任务，它跑完之后在频道里回复一整个线程，团队成员都能看到。之前这事得专人手动做，现在完全自动化了，至少省了半小时。唯一的问题是刚开始配置 Slack 连接的时候有点费劲，花了二十分钟才搞定 OAuth。

silverrabbit181

—

Rohit Prasad 离开亚马逊后第一个项目就是这个，看来他选开源这条路的决心挺大的。一个做过 Alexa 首席科学家的人跑来搞桌面 Agent，有点意思。

Jennifer_Wood_202112

—

MIT 许可证真的是很良心了，可以商用、可以改、可以自己部署。对于 SaaS 团队来说，等于白捡一个 Agent 底座，自己魔改一下就能用。

CosmwasmWizLarses

—

个人觉得当前版本最有价值的是它的审批流程设计。不是那种简单的「是/否」弹窗，而是每个工具调用都有对应的风险分级，shell 命令还强制逐次询问。这思路值得所有桌面 Agent 学习。

Jeffrey.NguyenJr

—

同时支持 Kimi、GLM、DeepSeek 这些国产模型，说明团队是真的在走模型中立路线。对于国内开发者来说，可以用国产模型做底座，不用被海外 API 绑定。

JStephens

—

我在团队里安排了四个预设角色:运维值班、销售简报、产品反馈整理和行政日程。每个场景都试了一下，最流畅的是整理日程，最需要打磨的是跨 Jira 和 GitHub 查版本进度的场景，有时候模型理解任务会出差错。

KimberlyColemanIII

—

开了 auto 模式让 Agent 全权操作，结果它差点把某个测试分支删了……还好执行 shell 命令之前还是弹了确认。建议非技术用户别开 auto，用 interactive 模式最安全。

Kelly_Ortiz369

—

项目刚开源不到一周就有 59 次 commit，维护节奏相当快。对比了一下 OpenClaw 的项目活跃度，OpenWorker 有吴恩达团队背书，长期维护应该更有保障。

DHill520

—

试了接 MCP 协议的工具，配了一个自定义连接器，整个过程比预期顺利。文档里说明支持 MCP，实际用起来也确实可以，扩展性比我想的好。

PElon

—

设了一个每天早上 8 点自动跑晨报的定时任务，跑了两天都正常。每次跑完有完整日志，可以看到 Agent 每一步做了什么，出了问题也好排查。

Joshua_Ortiz

—

我第一天就装上了，正儿八经用了一个星期。说下真实感受吧。优点是思路确实对，「交付成品」这个定位比现在市面上所有聊天式 AI 都先进，省去了很多手工操作。但缺点也很明显，首先 Windows 版没有代码签名，安装就要折腾半天；其次它本质是一个 Python 本地服务加一个前端壳，对非开发者来说，要搞懂配置 API Key、调 MCP 连接器这些东西，门槛实在不低。另外连接器的稳定性也还有提升空间，我试了 Gmail 和 Google Calendar，偶尔会断连需要重新授权。总的来说我觉得它适合愿意折腾的开发者，但不适合想开箱即用的普通用户。希望 beta 期能把这些问题都修好吧。

AnthonyEnnis

—

第一天就下载试了，config 方面确实需要点技术基础，但跑起来之后效果惊人。OpenWorker 和 OpenClaw 的竞争感觉会是今年下半年桌面 Agent 领域最值得看的事情。

Barbara.BellII

—

帮我准备客户简报这个功能实测效果很好，它会先扫描本地 CRM 数据、查邮件往来记录、再结合 GitHub 上的项目进度，生成一个完整的一页纸。要是早半年有这个，我至少能省几十个小时。

OpenWorker

In-depth Report

User Reviews

Similar Products