Andrew Ng’s OpenWorker delivers finished work, not chat—offline
Andrew Ng has just released OpenWorker, a new open-source desktop agent that skips the chat loop and delivers finished work—documents, spreadsheets, Slack replies with real data, updated calendars—while keeping everything on your machine. Instead of prompting for answers, it asks for the outcome you want, breaks it into steps, and uses vetted tools to get there safely. The whole stack runs locally, from the Tauri-based GUI to the FastAPI Python server and a permission engine that treats approvals as a first-class concern.
A shift from chat to delivered outcomes
Most AI assistants stop at generating text or code; OpenWorker pushes further. It prompts users for an outcome—a polished report, a triaged inbox, a scheduled meeting—then builds a plan, executes across local files and connected apps, and checks in before consequential actions. The architecture is split into four layers, all running on-device: a Tauri 2 shell wrapping a React 18 UI, a Python 3.10+ FastAPI server, a capability layer with local tools (files, Git, shell, search), and a model router that supports native, OpenAI-compatible, and local providers.
Permission and privacy by design
OpenWorker treats permissions as a typed layer, not an afterthought. Every tool call is classified by risk—read-only, write-local, exec, or external—and five permission modes govern what happens next. Interactive mode asks before writes and commands; auto mode allows path-scoped actions; and unattended mode routes pending prompts to an inbox instead of interrupting. All secrets stay local, and model keys never enter the context or traces. The only optional cloud component is an OAuth broker for integrations.
Bring your own model—or use curated ones
There’s no OpenWorker inference service. Users paste an API key or point the app to a local runtime. The curated model list includes 30 entries: OpenAI’s GPT-5.6 family, Anthropic’s Claude line, Google’s Gemini 3.1/3.6, OpenAI-compatible vendors like DeepSeek and Grok, and open-weight models via Together AI, Fireworks, and Ollama. The ops persona also treats content from tools, logs, and messages as untrusted, embedding a prompt-injection posture into the shipped agent.
Why it matters
OpenWorker signals a pivot from conversational AI to delivered outcomes—especially valuable for professionals who need reliable, private automation. By running locally and enforcing strict permission controls, it addresses privacy and security concerns that cloud-first agents often sidestep. For developers and teams, the open-source stack lowers the barrier to building specialized, trustworthy desktop agents without reinventing the wheel.
Source: MarkTechPost. AI-assisted editorial synthesis — TechnoExpress.