NemoClaw

NemoClaw is NVIDIA’s Apache-2.0 reference stack for running supported AI agents inside NVIDIA OpenShell sandboxes. OpenClaw is the default express-install choice, while current documentation also provides supported quickstarts for Hermes and LangChain Deep Agents Code. The stack combines an execution boundary, lifecycle operations, configurable inference, credential handling, and network policy. NVIDIA labels NemoClaw an alpha project, so it should be evaluated as developer and operator software rather than treated as a finished consumer assistant or a blanket security guarantee.

The documented workflow covers prerequisites, onboarding, sandbox creation, connection, monitoring, snapshots, backup, and restore. Operators can select a supported agent, configure local or remote inference, and review or modify egress rules for requests leaving the sandbox. Security still depends on host configuration, policy quality, mounted data, credentials, and the behavior of software inside the environment. Hardware, cloud GPU, model endpoint, and related NVIDIA service costs are separate from the open-source stack.

Top features

• Supported agent paths: run OpenClaw by default or follow documented setup paths for Hermes and LangChain Deep Agents Code.
• OpenShell sandboxing: place the selected agent inside a managed execution boundary instead of running it directly on the host.
• Lifecycle commands: create, connect to, inspect, monitor, snapshot, back up, and restore sandboxed environments.
• Network policy: inspect, approve, deny, and customize outbound access rather than granting unrestricted egress by default.
• Inference configuration: choose documented local or remote inference routes for a deployment.
• Credential guidance: follow current security procedures for storing and delivering secrets used by the sandboxed agent.
• Managed integrations and recovery: configure supported integrations and preserve environment state for testing or rollback.

Use cases

• Evaluate OpenClaw, Hermes, or LangChain Deep Agents Code inside a controlled sandbox before granting sensitive tools or data.
• Test outbound network policy and record the destinations an assistant actually needs.
• Compare local and remote inference routes while keeping the agent runtime separated from the host.
• Snapshot and restore an experimental environment during security, integration, or operations testing.
• Run a bounded proof of concept on supported infrastructure without presenting alpha software as production-certified.