AI scholar Andrew Ng released OpenWorker on July 23, introducing an open-source AI agent built to deliver finished work rather than simply talk with users.
According to the project description, OpenWorker can produce a completed document, send a Slack message, or update a calendar. Before it takes any important action, it asks the user to confirm.
Built around deliverables instead of prompts
OpenWorker is designed for users to specify the result they want, not just enter a single prompt. Examples in the article include helping prepare a client presentation, cleaning up a messy calendar, drafting a report, or doing an initial sort and judgment on Slack messages.
To handle that work, OpenWorker breaks a goal into multiple steps and executes across a user’s local files and everyday tools to produce an actual deliverable. If an action has consequences, the system returns to the user for approval before going ahead.
Local-first design with support for multiple model providers
OpenWorker uses a local-first architecture. Aside from services the user explicitly chooses to integrate, data stays on the local device.
Users can also bring their own API keys and pick from different model providers, including OpenAI, Anthropic Claude, Google Gemini, as well as Kimi, GLM, and DeepSeek open-weight models. It can also run locally through Ollama.
More than 25 integrations, with Mac first
Its GitHub project page says OpenWorker currently supports more than 25 integrations, including GitHub, Slack, Notion, Gmail, and Google Calendar, allowing the agent to complete tasks across tools.
OpenWorker is open-sourced under the MIT license. The code is available on GitHub at github.com/andrewyng/openworker, and the project also has a website at openworker.com. It currently runs on Mac, while a Windows version will be released later.
The article also noted that this deliverable-focused agent design lines up with a direction seen in other products. As one example, Anthropic previously launched Claude Cowork, which lets users teach Claude to complete specific work by recording their screen.