nemoclaw

NemoClaw is an open-source reference stack from NVIDIA for running supported AI agents more safely inside NVIDIA OpenShell sandboxes. It is aimed at users who want agent execution isolated from the host while still getting managed inference and controlled network access. Supported agents are OpenClaw (the default), Hermes, and LangChain Deep Agents Code. Each has its own quickstart guide, and the installer can either use express preset settings for a supported DGX or WSL host or walk through interactive choices such as agent, sandbox name, inference provider and model. Key capabilities described in the README: - Guided onboarding through the NemoClaw CLI and agent-specific aliases, plus a starter prompt for local coding agents like Cursor, Claude Code, Codex or Copilot. - Managed inference with selectable providers and routed inference configuration. - Network policy with baseline rules, operator approval flow, egress control, and static or dynamic customization with presets. - Sandbox lifecycle operations, snapshots, and managed integrations. - Sandbox hardening measures such as container security controls, capability drops and process limits. - Security best practices, risk framework and posture profiles. Documentation covers overview, architecture, ecosystem, prerequisites, inference provider selection, network policies, security best practices, sandbox hardening, CLI command reference and troubleshooting. The project is described as alpha, with maintainers reviewing issues and pull requests on a best-effort basis. Contributors can prepare a source checkout without creating a runtime sandbox using a dev-setup script, with optional flags for exposing a host-visible development CLI or enabling runtime sandbox validation. Security vulnerabilities should be reported through private NVIDIA channels rather than public issues. The project is licensed under Apache 2.0.