Andrew Ng Launches OpenWorker AI Desktop Coworker

Andrew Ng announced the release of OpenWorker on March 18, 2024, an open-source, local-first desktop AI agent designed to produce finished deliverables rather than engaging in conversational chat. Unlike traditional AI assistants that respond to prompts, OpenWorker is given a desired outcome, such as a polished document, a Slack message with specific data, an updated calendar, or a triaged inbox. The agent then autonomously breaks down the task into actionable steps, interacts with local files and connected applications, and seeks user confirmation before executing significant actions. This architecture ensures that all processing occurs on the user's machine, enhancing privacy and control.

The OpenWorker system is structured into four distinct layers, all operating locally. The desktop shell utilizes Tauri 2 for a native window that encapsulates a React 18 user interface, with the shell responsible for managing the Python server. The local agent server is built using Python 3.10+ with FastAPI and uvicorn, defaulting to communication on 127.0.0.1:8765. The configuration allows for a maximum of 12 model tool iterations per turn. The capability and connector layer integrates vetted local tools like file management, Git, ripgrep-based search, shell access, and to-do list functionality, alongside hosted integrations and MCP (Model Communication Protocol).

The model router layer provides a unified interface to various AI model providers, including native, OpenAI-compatible, reseller, and local runtimes. The core engine is powered by aisuite, Andrew Ng's provider-agnostic LLM library. OpenWorker does not host its own inference service; users are required to provide their API keys or point the application to a local AI model runtime. The curated model matrix includes 30 entries, featuring native support for OpenAI's GPT-5.6 (Sol/Terra/Luna) and GPT-5.5, Anthropic's Claude Fable 5, Opus 4.8, Sonnet 4.6, and Haiku 4.5, as well as Google's Gemini 3.1 Pro, 3.6 Flash, 2.5 Pro, and 2.5 Flash. Additionally, OpenAI-compatible vendors such as GLM-5.2, DeepSeek V4, Kimi K2.6, MiniMax M2.5, Qwen3 Max, and Grok 4 are supported.