OpenWorker

OpenWorker is an open-source, local-first desktop AI coworker from Andrew Ng, founder of DeepLearning.AI and AI Fund. Instead of returning chat, it returns finished work: a polished document, a filled spreadsheet, a drafted report, a Slack reply with the numbers, an updated calendar, a triaged inbox, or a security fix ready as a pull request. You describe an outcome ("prepare a customer brief," "untangle my calendar," "check where the release stands across Jira and GitHub"), and OpenWorker breaks it into steps, works across your desktop, files, and connected apps, and hands back the deliverable. Specialist coworkers arrive set up for one job: the Security Reviewer drives semgrep and secret scans and triages findings by reachability, Cloud Posture checks Terraform against a read-only view of AWS, and Dependency Audit proposes the smallest safe upgrade, each landing its fixes as pull requests you review. Sales, Executive Assistant, Marketing, and Ops On-call coworkers cover everyday work, and you can describe a new coworker in plain language.

Integration depth is a core strength. OpenWorker ships with 25+ native connectors, including Slack, GitHub, Gmail, Google Calendar, Google Drive, Microsoft Outlook, Jira, Confluence, Linear, Notion, monday.com, Asana, HubSpot, Attio, Zendesk, Stripe, Dropbox, and Box, plus direct access to your terminal and local files. Slack and GitHub are two-way: mention @OpenWorker in a Slack channel and a session opens on your desktop, the work runs with your tools, and the answer returns as a thread reply. Connections can be made in one click through an optional free OpenWorker account, which only brokers the OAuth handshake and passes tokens straight to your machine, or manually with your own tokens. Any tool reachable over the Model Context Protocol (MCP) plugs in as well, through a guided add-server flow with per-tool control. OpenWorker acts as an MCP client, consuming external MCP servers, and does not publish a documented MCP server of its own. Native SSO, SCIM, and centralized team administration are not part of the product, since it is a local, single-user desktop install rather than a managed cloud service.

Pricing is simple: OpenWorker is free and open-source under the MIT license, with no paid tiers for the software or for the optional account. It is model-agnostic and bring-your-own-key: you supply an API key for OpenAI, Anthropic, Google Gemini, DeepSeek, Qwen, Mistral, xAI Grok, Moonshot Kimi, MiniMax, Z.ai GLM, BytePlus Ark, or Thinking Machines Inkling, use open-weight models via Together or Fireworks, or run fully local and free with Ollama. Model usage is billed by whichever provider you choose, not by OpenWorker, so the effective cost is your own token spend, and running locally with Ollama has no per-token cost at all.

Because the app is free, the thing to understand is what free does not include. There is no hosted version of the agent and no managed model: you provide your own model key or run Ollama locally, and you pay the model provider directly for anything beyond a local model. The terms state that there is no service-level commitment, there is no enterprise support contract or audited compliance program, and the software is in open beta.

OpenWorker is best for developers, security teams, and knowledge workers who want a governed, local-first agent they can audit and run with the model of their choice. Organizations standardized on Microsoft 365 that want a cloud-hosted agent running inside their own tenant, with admin-selectable model routing, should consider Microsoft Copilot Cowork. People who want a personal assistant that lives in their text messages rather than a desktop coworker producing files should consider Poke ($16/mo billed annually).

Current state Q4 2026: OpenWorker launched on July 23, 2026 as an open beta and is developed in the open on GitHub (andrewyng/openworker, MIT license, roughly 18.6k stars). Version 0.3.0 on September 30, 2026 added a sandbox for what agents run. With NVIDIA OpenShell, each agent's shell and file tools run in their own Linux container that holds only the session's folders and an allow-listed set of network hosts. On a Mac the built-in macOS sandbox does the same with nothing to install, and on Windows a hidden local account does. The same release added Machines, which runs OpenWorker on another computer so sessions keep going when the laptop is closed, an openworker command line with a one-command Linux install, and an app interface in English and Simplified Chinese. Version 0.3.1 on October 5 published the command line as a container image for running a coworker inside OpenShell. The signed and notarized macOS app runs on Apple Silicon and Intel. Windows 10/11 installers ship with each release but are not yet code-signed, so SmartScreen warns during install, and the website still lists Windows as coming soon. There is no G2, Capterra or Product Hunt presence yet, and no audited security certifications such as SOC 2 or ISO 27001. The engine is built on Andrew Ng's open-source aisuite library.