Andrew Ng Launches OpenWorker: An Open-Source AI Agent That Runs Locally to Prevent Data Leaks
Discover OpenWorker, the new AI tool from Andrew Ng that goes beyond chatting to deliver completed work, featuring the freedom to choose models and Local-First data security.
Stock photo for illustration only, not from the actual event
- OpenWorker is an open-source AI agent designed to deliver actual deliverables, not just conversations.
- It operates on a Local-First basis on user devices, supporting integrations with over 25 services such as Slack, GitHub, and Notion.
- Users can freely choose various models via API or run them locally via Ollama without needing to connect to the cloud.
- It features a 4-tier risk assessment system to request user permission before executing critical actions, preventing data leaks.
When it comes to using artificial intelligence today, many people are accustomed to receiving lengthy recommendations and then having to type, format, and send them on their own. Recently, however, Andrew Ng, a renowned AI researcher and co-founder of Coursera, revolutionized the way we work by launching OpenWorker via his personal X account on July 23, 2026. He stated that this is an AI agent that doesn't just chat, but delivers fully completed work back to the user instantly, whether it's neatly formatted documents, Slack messages, or calendar organization.
The concept behind this tool is to shift away from traditional prompt-based chatting toward specifying a desired destination, such as a calculated spreadsheet or a reply message with actual figures. The AI then breaks down the goal into subtasks, operates across applications and local files, and returns the completed work. Use cases range from sales, marketing, and executive assistants to system administration involving troubleshooting.
The launch of OpenWorker reflects the direction of the AI industry as it transitions from conversational chatbots (Conversational AI) to action-oriented automated systems (Action-Oriented Agents). Andrew Ng's emphasis on a Local-First architecture and vendor independence addresses enterprise data privacy concerns that typically arise when confidential information must be sent to public clouds for processing.
Another key highlight is its flexibility in model selection. Because OpenWorker does not lock users into a single ecosystem, users can bring their own API keys—whether for cloud models like GPT-5.6 Sol, Claude Fable, and Gemini 3.6, or open-weight models such as DeepSeek, Qwen, Kimi, GLM, and Mistral. It also supports running fully offline via Ollama directly on the user's device without requiring API keys.
"Not just talking to you, but delivering completed work."
Andrew Ng
In terms of security, OpenWorker features a pre-execution permission system divided into four risk levels: read for side-effect-free data reading, write_local for modifying files within specified folders, exec for running commands, and external for sending data outside the device, such as posting to Slack. It also includes prompt injection defenses, treating data from tools or websites as untrusted.
The software is built with a desktop architecture utilizing Tauri 2 and React wrapping a Python server, leveraging aisuite as the foundation for model calls. OpenWorker is currently available for download in Open Beta on macOS under the MIT license with no software fees. Interested individuals can find further details and source code on the official website.
Source: Techsauce
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment