Ng's OpenWorker: AI That Delivers

Andrew Ng, a pivotal figure in AI, has unveiled OpenWorker, an open-source agent aimed at transforming desktop productivity. Unlike conventional chatbots, OpenWorker is engineered to deliver tangible outputs: polished documents, calendar updates, or triaged Slack messages.

A Desktop-First AI Co-Worker

The OpenWorker AI agent runs natively on macOS, with Windows support imminent. This local-first approach ensures user data remains on the machine, only interacting with chosen LLM providers and integrations.

Users can connect their own API keys, supporting a range of models from GPT-5.6 Sol and Claude Fable to open-weight options like Kimi and GLM, or even Ollama for fully local processing.

OpenWorker tackles tasks such as preparing customer briefs, managing calendars, or drafting reports. It breaks down complex requests into actionable steps, working across local files and over 25 integrated tools, including GitHub, Slack, Jira, and Google Calendar.

Crucially, the system incorporates a human-in-the-loop mechanism. Before executing any impactful action, like sending a message or altering a calendar entry, OpenWorker seeks user approval. This design emphasizes control and prevents unintended consequences.

Privacy and Flexibility at its Core

Ng and co-creator Rohit Prasad developed OpenWorker to provide an open, privacy-preserving, and model-independent solution for AI automation for desktop. The agent's core engine, conversation history, and API keys are stored locally, with a minimal cloud service for OAuth handshakes.

This commitment to user privacy extends to its operational model. OpenWorker's ability to automate workflows with scheduled tasks, like daily briefs or weekly reports, also includes an approval inbox for unattended runs, ensuring oversight.

The project, currently in open beta, is built on aisuite, a lightweight Python library for LLM integration and agent development. Its source code is publicly available on GitHub, inviting community contributions and scrutiny.