Andrew Ng has released OpenWorker, an open-source desktop agent that hands you finished work instead of a chat transcript. Announced on July 23, 2026 under an MIT license, it drafts documents, sends Slack messages, and updates calendars locally on your Mac using your own model API keys.
How to Try It
OpenWorker is a builder's tool you can run today. Clone the repository on GitHub, add an API key for OpenAI, Anthropic, or Google, and point it at your files, Slack, and calendar. Prefer to keep everything offline? Connect a local model through Ollama and no cloud key is required. Ask it to prepare a customer brief or triage a Slack thread, and it produces the actual deliverable while pausing for your approval before anything consequential.
Why It Matters
Most desktop agents bolt an approval prompt onto the interface. OpenWorker instead classifies every tool call into one of four typed risk classes before it runs: read (no side effects, always allowed), write_local (path-scoped file changes), exec (runs commands), and external (off-machine actions). That gives creators building automations a predictable safety model rather than a blanket "allow" button. It joins a growing wave of open, local-first agent workspaces, alongside projects like Buzz Block, aimed at builders who want control over where their data and models run.
Key Details
Model access: Bring-your-own-key across 30 curated tool-calling models, built on Ng's provider-agnostic aisuite library, plus fully local Ollama.
Stack: A Tauri 2 and React desktop shell over a local Python FastAPI server. The initial release ships roughly 32,000 lines of Python and 78 backend test modules.
Connectors: Around 35 integrations for files and everyday tools, with data staying on your machine except through the providers you choose.
Platform: Mac at launch, with Windows support announced as coming soon.
What to Do Next
Read the project overview at openworker.com to see the full connector list and setup steps, then decide whether a local, model-independent coworker fits your workflow before wiring it into production tasks.
