Andrew Ng releases OpenWorker open-source agent that avoids vendor lock-in

Andrew Ng, known for his work at Google Brain and Coursera, has launched an open-source beta of OpenWorker under an MIT license. The project provides a code-first agent that breaks a task into steps, accesses files, the terminal and more than 25 external services, without tying users to a single model.

The user specifies a desired result—a summarized document, an organized inbox, an updated calendar entry, or a Slack reply with the correct numbers—and the agent decides how to achieve it. Integrated services include GitHub, Jira, Notion, Gmail, Google Calendar, Outlook and HubSpot, allowing the tool to interact with most channels where organizational data flows. OpenWorker does not include its own model; users supply an API key for OpenAI, Anthropic, Google, DeepSeek, Qwen, Mistral, GLM, Kimi, Grok or run a model locally via Ollama. The code is the orchestra, not the player.

The MIT license permits commercial use, modification and redistribution, but the weights of any models run on the platform remain under their respective licenses. Ng stated on X that the aim is to give the market an alternative that does not depend on a specific vendor, a claim reinforced by code that can be hosted on a private server without sending data externally.

The macOS build is digitally signed and updates automatically. The Windows build is still awaiting signing, so SmartScreen will issue a warning on the first run—a common hurdle for anyone distributing binaries without an EV certificate. The team says the Windows signature will be added later; meanwhile users can bypass the warning by manual execution or by running the source directly.

The combination of Ng’s broad reach (his courses have reached more than 8 million learners), a permissive license and a model-agnostic architecture creates a rare starting point: an agent infrastructure that organizations can adopt without signing a SaaS contract or betting on the survival of any single startup. The code is on GitHub, the license is clear, and the choice of model stays in the user’s hands.