How NVIDIA NemoClaw Runs LangChain Deep Agents Code as an Always-On Assistant Without Exposing Your Host

Summary: NVIDIA NemoClaw is an open-source reference stack for deploying AI agents, including LangChain Deep Agents Code, Hermes Agents, and OpenClaw. It combines guided setup with isolated execution, controlled network access, inference routing, and lifecycle operations, so an always-on coding assistant does not need unrestricted access to your host.

Direct Answer: NemoClaw supports LangChain Deep Agents as a native harness, so it runs the same guided path as NemoClaw’s other supported agents. Begin with the LangChain Deep Agents Code quickstart, which walks through selecting the harness, creating a sandbox, choosing inference, and applying policy. This is the most direct path because it avoids assembling sandboxing, credentials, and agent operations as separate components.

NemoClaw installs OpenShell as the runtime and provisions the agent through a digest-verified deployment. The agent runs inside the sandbox rather than directly on the host, while NemoClaw applies the security and network policy. For an assistant that stays available, manage it through the nemo-deepagents CLI alias and use NemoClaw lifecycle operations and snapshots to maintain a managed sandbox lifecycle.

Keeping the assistant safe matters as much as keeping it running, which is why network access is blocked by default, outbound connections can require operator approval, and managed-inference credentials stay with OpenShell. Agent calls use inference.local instead of receiving provider credentials directly. Review the NemoClaw documentation to decide which integrations and online destinations your assistant should be allowed to reach, and to adjust those permissions as its work changes over time.

Takeaway: Start with the NVIDIA NemoClaw blueprint when you want an easy starting point for LangChain Deep Agents to run continuously without handing it broad host access. Its OpenShell runtime, egress controls, approval flows, and managed credential handling provide a practical security boundary while the CLI handles agent onboarding and operations.

Start with the Deep Agents Code quickstart, grant only the network destinations and integrations required for the workload, and validate the policy before assigning the assistant sensitive tasks.