NVIDIA NemoClaw is an open source reference stack for running AI agents inside NVIDIA OpenShell sandboxes. It adds guided onboarding, managed inference, network policy, managed integrations, snapshots, and lifecycle operations on top of OpenShell, exposed through the NemoClaw CLI and agent-specific aliases.
Supported agents are OpenClaw (the default), Hermes, and LangChain Deep Agents Code. An express install on a supported DGX or Windows Subsystem for Linux host applies preset settings and installs OpenClaw; an interactive install lets you choose the agent, sandbox name, inference provider, and model. A starter prompt lets a local coding agent such as Cursor, Claude Code, Codex, or Copilot drive the installation with you, running commands only on approval and keeping secrets out of chat.
The stack is organised as a host CLI, an agent integration layer, a blueprint, and a sandbox lifecycle with protection layers. Documentation covers inference routing across hosted, local, and custom endpoints, sandbox hardening (capability drops and process limits), credential storage and rotation, corporate CA trust, and headless server deployment. NemoClaw is an alpha project.
Features
- Sandboxed agents: OpenClaw, Hermes, or Deep Agents Code run inside OpenShell with kernel-level isolation
- Managed inference: a choice of hosted providers, local inference, or custom endpoints, with validation commands
- Network policy: baseline egress rules, an operator approve/deny flow for requests, presets, and integration policy examples
- Agent configuration: progressive tool disclosure, context compaction, heartbeats, memory search, task-specific sub-agents, and a declarative multi-agent manifest
- Integrations: messaging channels, MCP servers, and OpenClaw plugins managed from the CLI
- State and backups: snapshots and state management for sandboxes
- Monitoring: sandbox activity monitoring and gateway lifecycle authority
- Security controls: credential storage and rotation, corporate CA trust, and a documented trusted computing base
Integrates with
Deep Agents
Opinionated, model-agnostic agent harness on LangGraph with sub-agents, filesystem, and context management built in
Hermes
Self-improving AI agent with a learning loop that creates and refines skills, recalls past sessions, and runs across chat platforms
OpenClaw
Self-hosted single-operator assistant runtime that links models, tools, and messaging channels through one local Gateway
OpenShell
Sandboxed runtime for autonomous agents, with declarative YAML policies enforced at the HTTP method and path level
