Andrew Ng releases OpenWorker, an open-source desktop agent
Andrew Ng shipped an MIT-licensed, local-first desktop agent that plans tasks across files and apps and hands back finished deliverables instead of chat.
Andrew Ng has released OpenWorker, an MIT-licensed desktop agent co-built with Rohit Prasad that stakes out a distinct position in the crowded agent market: open source, local-first, and judged on finished deliverables rather than conversation. Ask it for an outcome — a polished document, a triaged inbox, an updated calendar — and it decomposes the goal into steps, works across local files and more than 25 integrations including Slack, email, calendar and MCP servers, checks in before any consequential action, and hands back the completed artifact.
Local-first by architecture
OpenWorker runs on the user's own machine — macOS today, Windows promised — as a Tauri desktop shell in front of a local Python agent server. There is no hosted backend: users bring their own API keys, and data leaves the computer only through whichever model provider or integration the user selects. A provider-agnostic router supports roughly 30 curated models spanning GPT-5.6, Claude, Gemini, DeepSeek, Mistral and Grok, open-weight models like Kimi and GLM through Together and Fireworks, and fully local inference via Ollama. The project builds on aisuite, Ng's earlier multi-provider library, and drew nearly a thousand GitHub stars within hours under an open-beta label.
The counter-positioning
The framing — delivers finished work instead of chat — reads as a direct answer to the proprietary desktop coworkers now shipping from the big labs, Anthropic's Claude Cowork and OpenAI's ChatGPT Work among them. Those products bundle the agent harness, the model and the data pathway into one closed subscription. OpenWorker unbundles all three: the harness is open code, the model is swappable, and the data stays home.
Why it matters: the agent harness is emerging as this year's key competitive layer, and the big labs are betting it will be proprietary. A credible open, model-independent alternative from one of the field's most trusted educators applies commoditization pressure at exactly that layer — and gives enterprises with privacy constraints a viable path to desktop agents that never ship their files to a vendor cloud. Whether a community project can match the polish of funded products is now an empirical question with a public repo attached.