Andrew Ng releases OpenWorker, an open-source desktop AI agent

The agent runs locally and hooks into the file system and terminal directly, producing actual deliverables: documents, spreadsheets, reports, calendar updates, cleaned-up inboxes. Anything that could cause real damage — sending a message, editing a calendar, running a shell command — sits behind an approval gate that has to be cleared before anything happens, a distinction from agents that just execute and report back after the fact.

Connectivity spans more than 25 integrations, including GitHub, Jira, Notion, Linear, HubSpot, Outlook, Gmail, Google Calendar, and monday.com, plus MCP-compatible tools for anything not natively supported. A Slack integration lets anyone trigger a session just by mentioning @OpenWorker in a channel, with the agent replying in a thread once the work is done. Scheduled automations handle recurring jobs on their own — things like a daily morning brief or a weekly report that would otherwise need manually re-running.

Users aren’t locked into one model provider. Cloud options include OpenAI, Anthropic, Google Gemini, DeepSeek, Mistral, Grok, Kimi, Qwen, and MiniMax, alongside open-weight models routed through Together and Fireworks. Anyone who wants the whole thing running without a cloud dependency can point it at Ollama instead and keep every model call on their own hardware, with no data leaving the machine.

That local-first approach carries through to storage too — agent loops, conversation history, connector tokens, and API keys stay on the machine itself, with a small cloud component handling only OAuth logins during sign-in. Installers are available for Apple Silicon Macs running macOS 12 or later and Windows 10/11 x64, and the whole thing is built on aisuite, Ng’s own library for talking to different LLM providers through one interface.