Skip to content
Python
2.3 KB
Raw Blame

Build your product. OpenWorker handles the agents.

Creating an app that uses AI should be easy. Builders should be able to focus on a useful, thoughtful product experience while the harness handles agent loops, permissions and alignment. OpenWorker Harness SDK gives you the foundation for Claude- or Codex-style agent experiences—conversation, tools, streamed activity and ongoing sessions—inside your own product.

Build a desktop app for your users, a web app deployed to your cloud, or both. The Rust and TypeScript SDKs let you choose the deployment that fits your product. Guardrails, sandbox integration, an optional auto-approver, session durability, agent alignment controls, service scaling and support for multiple LLM providers are built into the platform, so you can configure these capabilities instead of implementing them from scratch.

You can configure NVIDIA OpenShell to isolate your agents' tool execution. The harness coordinates model calls, policy and sessions; the sandbox bounds what its ToolRunner can access. See OpenShell integration for that boundary.

Choose a track

Both tracks include a small UI, agent creation and packaging/deployment instructions. Open the source drawer to explore the complete example files alongside the guide.

       Desktop app                 Web application
       Your interface              Your API + browser UI
              │                          │
              └────────────┬─────────────┘
               OpenWorker Harness SDK
                           │
                  OpenWorker Harness
           Models · Tools · Sessions · Teams
          Guardrails · Approvals · Observability
                           │
           Tool execution: local or OpenShell

Start with the example, then choose the models, tools and deployment your product needs. Read versions and compatibility for package availability and safe application upgrades.