Andrew Ng's OpenWorker wants to do your work, not just chat about it

Andrew Ng, the AI educator who co-founded Coursera and once led Google Brain, released a free desktop program on Thursday that tries to finish office tasks across the apps people already use, rather than just answer questions about them.

The program, OpenWorker, is published under the permissive MIT license on GitHub. Ng announced it on his X account on July 23, and Chinese AI outlet QbitAI published a feature walkthrough the same day. MarkTechPost followed with corroborating coverage.

Several competing desktop and cloud agents have shipped this quarter, including Manus, Cognition's Devin, and OpenAI's Operator. OpenWorker's differentiator is the four-bet combination it ships with.

Four bets, one agent

OpenWorker stacks four choices that most prior desktop agents split up:

Open source, under MIT, so the agent loop and connector code can be read, audited, and forked.

Local-first by default. The agent loop, conversation history, and the tokens that talk to third-party apps stay on the user's machine. Hosting an LLM or sending a request to a third-party tool is an explicit opt-in.

Model-agnostic. The user supplies the model and the API key. Per the project's documentation, OpenWorker works with hosted models including OpenAI's GPT 5.6 Sol, Claude Fable, and Google's Gemini 3.6, plus open-weight models like Kimi, GLM, DeepSeek, and Inkling served locally through Ollama.

A confirmation gate. Before the agent sends a message, modifies a calendar, writes to an external tool, or runs a terminal command, it pauses and asks.

The confirmation gate is the bet most prior open-source desktop agents have skipped. Earlier projects, including Open Interpreter, have generally defaulted to executing commands in the foreground without a confirmation step, and Anthropic's computer-use work has handed screen control to the model with a similar default. OpenWorker's check-in-first design treats the user more like a manager reviewing a junior coworker's draft than an absent principal.

What it actually connects to

OpenWorker ships with connectors for more than 25 third-party tools, including GitHub, Slack, Jira, Notion, Linear, HubSpot, Outlook, Gmail, and Google Calendar. The connector set extends through MCP, the Model Context Protocol, an emerging standard that lets the agent plug in additional tools at runtime. The connector tokens are stored on the device by default, which is what makes the local-first claim meaningful for teams that already route their work through Slack and Jira.

The agent also supports scheduled jobs in coordination with an LLM, so a user can ask it to run a daily briefing, a weekly report, or a Slack channel monitor without launching it each time. The same confirmation gate applies. If the user is offline, OpenWorker queues the request and waits.

What the design pattern is actually worth

The four-bet stack is the genuine news. Each bet has been tried before; the combination has not.

The confirmation gate does most of the work, and it is also the most fragile. The gate is only as useful as the away-mode queue. If the queue is too quiet, the user comes back to a backlog of decisions. If it is too loud, the user disables the gate. The source coverage does not describe the queue in any detail, and that is the piece to test in the next week.

Local-first is conditional. The data stays on the device, but the model calls do not, unless the user runs a local model through Ollama. The "open source" label covers the agent runtime and connectors; it does not cover the model itself, which is whatever the user has paid for elsewhere.

The "finished artifact" framing is closer to the truth than the typical "AI coworker" pitch. The agent's job, per the GitHub README, is to return a polished document, a sent Slack message, or an updated calendar entry, not a chat transcript. Whether that framing holds depends on the model the user brings.

Where it fits

OpenWorker lands in a week of competing agent announcements, including Manus, Cognition's Devin, and OpenAI's Operator. Ng's differentiator is the open + local + model-agnostic + confirmation-gate combination, paired with the distribution that comes from a former Google Brain lead and a Stanford adjunct who built the DeepLearning.AI and Coursera audiences.

The binary runs on macOS only. Ng said a Windows version is coming. The GitHub repository is the source of truth for capability claims, and that is where to look for what the next commit does to the confirmation gate.