Skip to main content

Andrew Ng Open-Sources OpenWorker: A Local-First Paradigm for Desktop AI Agents

An AI That Does Real Work — on Your Own Computer​

On July 24, 2026, Andrew Ng posted a brief announcement on LinkedIn: he was open-sourcing a desktop AI agent called OpenWorker. Within 48 hours, the project had crossed 3,700 stars on GitHub.

This is not the first desktop AI agent. But OpenWorker stands out because it explicitly embraces and executes a design philosophy that is the polar opposite of mainstream cloud-based agents: local-first, model-agnostic, and human-in-the-loop. At a moment when the AI industry is debating whether agents should run in the cloud or on-device, Andrew Ng cast his vote — in code.

What Is OpenWorker? An AI Coworker That Lives on Your Desktop​

OpenWorker's positioning is crystal clear. It is not a chatbot. It is an AI agent that delivers finished work. The official site illustrates several typical scenarios: in sales, it can compile client backgrounds, spot expansion signals, and prepare follow-up materials for review; in an executive assistant role, it can organize calendars and distill tasks and decisions from email threads; in marketing, it can aggregate campaign signals, explain shifts, and compile reports; in operations, it can correlate alerts with recent changes and runbooks.

The workflow is straightforward: the user describes the goal in plain language — "I'm renewing with Acme tomorrow at 10 AM, help me pull together the materials I need to review" — and OpenWorker coordinates execution locally: pulling customer data from HubSpot, finding correspondence in email, reading related documents from the file system, then generating a structured renewal briefing. Only after the user reviews and approves does it send a follow-up email.

Throughout this process, all data stays on the user's machine. Credentials are stored in the local keychain. Model calls use the user's own API keys. For critical actions — sending emails, posting Slack messages, executing commands — OpenWorker pauses and waits for explicit approval.

Three Design Principles That Target Cloud Agent Pain Points​

OpenWorker's design philosophy can be summarized in three principles, each precisely addressing a core tension in the desktop agent space:

Principle 1: Local-first. OpenWorker runs locally by default. The user's files, emails, and calendar data never leave their computer. This directly addresses the number one concern enterprises have about AI agents: data security. When an AI agent needs access to your entire email history, calendar, and customer data to deliver value, uploading all of that to a third-party cloud server is a non-starter. Local-first is not a nice-to-have — it is a prerequisite for AI agents entering enterprise environments.

Principle 2: Model-agnostic. OpenWorker does not lock users into any single model. You can use OpenAI's GPT, Anthropic's Claude, Google's Gemini, or Chinese models like Kimi, GLM, DeepSeek, and Qwen — or run entirely local models through Ollama. The underlying architecture is built on aisuite, a provider-agnostic Python framework that Ng open-sourced earlier, offering a unified API across multiple large language model providers. Every task can use a different model, and you can switch mid-conversation.

Principle 3: Human-in-the-loop. This is the key design decision that sets OpenWorker apart from fully autonomous agents. Users not only set goals but also define tool scopes and approval rules. Sending emails, posting Slack messages, or running commands all require user sign-off before execution. This human-AI collaboration model lowers the psychological barrier — users know the AI will not do anything without their knowledge.

The Tech Stack: Lightweight but Complete​

OpenWorker's architecture is clean and practical. It achieves model independence through the aisuite library and integrates with over 25 tools — from Slack, Outlook, and Gmail to Notion, HubSpot, GitHub, Jira, and Linear — covering the primary productivity tools knowledge workers use daily.

Event-driven automation is another highlight. Users can configure trigger events: when mentioned on Slack, automatically compile incident context and recent changes; when a specific email arrives, prepare a draft reply for review; before a meeting, pull the latest data. But critical actions still require approval — the designers clearly understand that uncontrolled automation is more dangerous than no automation at all.

On the model access front, OpenWorker's supported list reads like a who's who: OpenAI, Anthropic, Google, xAI, Mistral, Together AI, and Fireworks on the international side, plus GLM, Kimi, DeepSeek, Qwen, and MiniMax from China, along with Ollama for fully local deployment. This breadth of model support is among the best in the desktop agent category today.

The macOS version is officially available, and the Windows version is functional but not yet code-signed. The project uses the MIT license, is 100 percent open source, and lives at github.com/andrewyng/openworker.

The Desktop Agent Paradigm War​

OpenWorker's release intensifies the cloud-versus-local debate in the desktop AI agent space.

On one side is the cloud-agent approach favored by certain major AI providers — the agent's core logic runs in the cloud, and the desktop client is merely a frontend interface. The advantage is that computational power is not constrained by local hardware, and product iteration is faster. The cost is that user data must leave the local device.

On the other side is the local-first approach represented by OpenWorker — the agent's orchestration logic and data processing happen locally, with only inference calls going to cloud models via API. Data stays on-device, and users maintain full control over access permissions and approval workflows.

OpenWorker's choice of the latter is not just a technical preference — it is a product stance. Ng has spent years in AI education and deeply understands enterprise sensitivity around data security. His LinkedIn announcement did not discuss model parameter counts or inference speed. Instead, it emphasized: "Important actions are always approved by you." That sentence itself is a product manifesto.

Beyond 3.7k Stars: Why Developers Are Buying In​

OpenWorker's rapid accumulation of over 3,700 stars in under 48 hours is impressive for a desktop agent project. What drives this number is not just Ng's personal influence, but a genuine demand among developers for controllable AI agents.

Over the past year, the AI agent concept has been hyped relentlessly, but most products have hit the same wall at deployment: users do not trust them. They do not trust the agent to read their email, manipulate their files, or send messages on their behalf. These trust deficits all trace back to one root cause — users cannot see what the agent is doing, nor can they effectively control it.

OpenWorker solves this with three steps: data stays local, operations require approval, and the code is fully open-source and auditable. Together, these lower the trust barrier dramatically. A telling detail: community-contributed localization packs and custom API projects have already appeared on GitHub, indicating that developers are not just using OpenWorker — they are building an ecosystem around it.

Conclusion: The Next Phase of AI Agents — From Smarter to More Trustworthy​

The significance of OpenWorker is not that it implements a few more features than other desktop agents. It is that it articulates a complete design paradigm. This paradigm answers a question more fundamental than any technical implementation: what kind of AI agent would users actually trust?

Andrew Ng's answer: one that runs on my own computer, uses the model I choose, and asks for my approval before every important action. This answer is simple but precise. As AI agents graduate from toys to tools, from demos to productivity, trust will become a scarcer resource than intelligence. OpenWorker, with its 100 percent open-source and local-first design, has provided a blueprint for that path.