Guozhen AIGlobal AI field notes and model intelligence

Realtime AI News

Andrew Ng Releases OpenWorker: An Open-Source Desktop AI Agent That Delivers Finished Work, Not Chat Replies

Andrew Ng has released OpenWorker, a fully open-source desktop AI agent under the MIT license. Unlike traditional AI chatbots, OpenWorker focuses on delivering completed deliverables — a polished document, a Slack reply with real data, or an updated calendar — rather than conversational text outputs.

Published
吴恩达发布100%开源桌面AI Agent OpenWorker:交付成品而非聊天回复
Image source: marktechpost.com

Andrew Ng has officially released OpenWorker, a fully open-source desktop AI coworker licensed under MIT. Unlike conversational AI products that produce text replies, OpenWorker asks users for an outcome — a finished document, a Slack reply containing the actual numbers, or a triaged inbox — then breaks that outcome into steps and executes them across local files and connected apps.

The architecture consists of four layers, all running on the user's local machine. The desktop shell uses Tauri 2 wrapping a React 18 UI, which supervises a local Python agent server built on FastAPI and uvicorn (default binding 127.0.0.1:8765). The capability layer covers local files, git, ripgrep-backed search, shell, and todo tools, plus hosted integrations through MCP. A model router unifies calls across local and remote providers.

The engine is built on aisuite, Ng's provider-agnostic LLM library. OpenWorker offers no hosted inference — users paste their own API key or point the app at a local runtime. A curated model matrix of 30 entries covers providers including OpenAI, Anthropic, Google, DeepSeek, GLM, Kimi, MiniMax, Qwen, and Grok, plus fully local models via Ollama.

The permission engine is the standout engineering feature. Every tool call is classified into one of four risk classes: read-only, write_local (path-scoped), exec (command execution), and external (off-machine side effects). Five permission modes govern behavior, with the default interactive mode asking before writes, commands, and external actions. Unattended mode routes prompts to an inbox rather than raising the autonomy ceiling.

Privacy is designed as local-first by default. Model calls go directly from the machine to the configured provider. Conversations, connector tokens, and model keys stay local. The only cloud component is an optional OAuth broker that handles connector handshakes using Auth0 Authorization Code with PKCE, and connector tokens never leave the user's machine.

OpenWorker represents Ng's strategic bet on privacy-preserving, practical AI agents. It offers developers and enterprises a self-hosted, fully controllable desktop automation solution that could accelerate real-world integration of AI agents into personal workflows.

Why it matters

OpenWorker sets a new open-source benchmark for local-first, privacy-preserving desktop AI agents, giving developers a self-hosted alternative that could accelerate enterprise adoption of agentic automation.

Andrew NgOpenWorkerOpen SourceDesktop AgentAI Agent
Back to realtime news

Nearby Updates

All

07/25, 13:45

Kimi US Researchers Say China-US AI Gap Narrowed to 3-5 Months, Challenging OpenAI and Anthropic Valuations

Sina Finance reports that interviews with Kimi (Moonshot AI)'s researchers based in the United States reveal the China-US AI technology gap has narrowed to 3-5 months. This development is shaking the long-term valuation logic of OpenAI and Anthropic, as Chinese large language models increasingly disrupt Silicon Valley's competitive landscape.

07/25, 10:47

Nvidia, Meta, Microsoft and Other Tech Giants Urge US Government Not to Restrict Open-Weight AI Models

Multiple AI companies including Hugging Face, Meta, Microsoft, Mistral, and Nvidia have signed an open letter urging policymakers against imposing broad premature restrictions on open-weight AI models. The letter comes as Washington debates potential bans on Chinese open-weight models following allegations of intellectual property theft.

07/25, 09:20

Xiaomi Mysterious Flagship Clears All Three Certifications: Self-Developed Xuanjie Chip, OS, and AI Model

A mysterious Xiaomi flagship phone has obtained all three Chinese network access certifications, and will feature its self-developed Xuanjie chip, proprietary operating system, and in-house AI large model. This marks the first time Xiaomi has achieved full self-developed chip-OS-AI integration in a high-end device.

07/25, 08:23

OpenAI Launches AI Keypad: Coding Boost or Novelty Toy?

OpenAI has released a new AI keypad that early testers describe as highly engaging for coders but potentially mystifying for general users. The device embeds AI capabilities directly into a hardware input peripheral.