Create an autonomous coding agent, a support agent that investigates and resolves issues, or a team that coordinates a workflow across your systems. OpenWorker Harness SDK handles model calls, tool execution, delegation and durable sessions so you can focus on your product experience.
Choose a track
- Build a desktop app. Start with Code Atlas: select a code directory and stream an explanation. Then add tools and delegation for more capable workflows.
- Build a web app. Start with PDF Notes: upload a document and ask follow-up questions. Deploy the frontend and API together on Railway or your container platform.
The tutorials deliberately start small. The same SDK supports long-running work, multiple agents and remote service deployments as your application grows.
Add the capabilities your product needs
| Your product needs | Start here |
|---|---|
| Agents that act on your systems | Application tools and MCP |
| Work split between agents | Sub-agents and teams |
| Conversations that survive restarts | Sessions and recovery |
| Images, PDFs and text files | Attachments |
| Control over actions and sensitive content | Approvals · Guardrails and hooks |
| Visibility into latency, token usage and failures | Observability |
Choose cloud or local models. Run the Rust harness in-process in a Rust application, or as a managed child process through the TypeScript SDK. A browser talks to your application server; credentials stay there.
Understand the architecture
Use NVIDIA OpenShell to isolate configured tool execution. The harness coordinates the agent outside that sandbox; application callbacks retain your application’s privileges. See architecture for the process and storage boundaries.
Start inside your app. Move to a shared service when you need independent worker capacity and shared storage. Deployment options distinguishes application hosting from experimental cluster setup.
See versions and installation for published packages, Rust setup and runtime compatibility.