OpenWorker v0.2.0 adds security coworkers, skills

OpenWorker, the open-source desktop agent from Andrew Ng and Rohit Prasad, shipped v0.2.0 on 24 August 2026 with security-focused coworkers and new skills support. A follow-up release, v0.2.1, landed the next day adding another model option.

OpenWorker v0.2.0, released 24 August 2026, adds Skills for packaging reusable workflows, three specialized security coworkers for code review, cloud posture, and dependency audits, persistent cross-session memory, an auto-approve mode for routine tool calls, and Intel Mac builds. v0.2.1, released 25 August 2026, adds the Ox Alpha model via OpenRouter.

What changed

• Security coworkers — three specialists join the picker: a Security Coworker for code review, a Cloud Posture Coworker for infrastructure checks, and a Dependency Audit Coworker for supply-chain scans, hardened with locked-down secrets-file permissions and pinned web-fetch connections.
• Skills — reusable instruction packs the coworker loads on demand, alongside persistent cross-session memory, an auto-approve mode for routine tool calls, a guided MCP setup flow, new reasoning-effort levels and Ark model providers, and Intel Mac builds.
• v0.2.1, out 25 August 2026, adds Ox Alpha, a stealth-preview model via OpenRouter with a 1M-token context window.

Why it matters for agent builders

Security work favors local-first, model-agnostic agents: a dependency or cloud-posture scan can run entirely offline, with open-weight models via Ollama, so proprietary code never reaches a third-party API. That pitch sits alongside coding agents such as Claude Code and Cursor, both tracked on AI Agents Listing, which handle a different slice of the same audit-and-ship workflow. OpenWorker's MIT license also lets teams read the approval and tool-calling logic before trusting it with a shell.

What to watch

OpenWorker has no independent security benchmark yet, so claims about catching real vulnerabilities rest on Andrew Ng's framing and early user reports rather than published third-party evaluation. Windows builds also remain unsigned, which triggers a SmartScreen warning during install.

Key takeaways

• OpenWorker v0.2.0 (24 August 2026) adds Skills, three security coworkers, persistent memory, and Intel Mac builds.
• v0.2.1 (25 August 2026) adds the Ox Alpha model via OpenRouter with a 1M-token context window.
• OpenWorker is free and MIT-licensed; users bring their own model API key or run models locally with Ollama at no cost.
• The GitHub repository shows 15.3k stars, 2.1k forks, and 16 contributors as of 25 August 2026.
• Andrew Ng frames the release around cybersecurity workflows, calling OpenWorker an open, auditable option for defenders.

FAQ

Is OpenWorker free to use?

Yes. OpenWorker itself is free and MIT-licensed; the only cost is the model you connect it to, whether that's an API key for OpenAI, Anthropic, or another provider, or a local open-weight model through Ollama, which runs at no additional charge.

Does OpenWorker run on Windows?

Yes. A Windows 10/11 x64 build ships alongside the macOS app with the same coworker and connectors. The installer isn't code-signed yet, so SmartScreen shows an unknown-publisher warning during setup while signing is in progress, per the project's GitHub repository.

How is OpenWorker different from Claude Code or Cursor?

OpenWorker runs locally with any model you pick, including free local ones via Ollama, while Claude Code and Cursor, both listed on AI Agents Listing, tie to a specific vendor's cloud models. OpenWorker focuses on general desktop and security work, not a coding editor.

Source: OpenWorker v0.2.0 release notes Discussion: X post

Browse developer tools on AI Agents Listing