NemoClaw is a reference stack for running always-on AI agents inside OpenShell sandboxes. It installs OpenShell, applies an agent integration layer and a versioned blueprint for the chosen agent, and then manages that sandboxed agent through the NemoClaw CLI and its agent-specific aliases.
Supported agents are OpenClaw (the default), Hermes, and LangChain Deep Agents Code. Qualification-only candidate agents stay separate from the supported runtimes. Around the agent, NemoClaw adds guided onboarding, managed inference with a choice of provider and model, network policy with an operator approval flow for egress, managed integrations, OpenShell-backed state persistence, and lifecycle operations. OpenShell holds custody of the credentials for managed inference and host-configured integrations.
Setup runs either through a terminal installer with an express preset for supported DGX and WSL hosts, or through a starter prompt that a local coding agent follows: it asks one question at a time, runs commands only with approval, and keeps secrets out of chat. NemoClaw is an alpha project.
Features
- Supported agents: OpenClaw (default), Hermes, and LangChain Deep Agents Code, each run inside an OpenShell sandbox
- Guided onboarding: an interactive wizard picks the agent, sandbox name, inference provider, and model, with an express preset for supported DGX and WSL hosts
- Managed inference: routed inference configuration with provider validation, and credentials held by OpenShell rather than the agent
- Network policy: baseline egress rules, an operator approval flow, and static or dynamic policy changes with presets
- Sandbox hardening: container security measures, capability drops, and process limits
- Versioned blueprints: each agent integration applies a versioned blueprint that defines the sandbox lifecycle
- State persistence: agent state persists through OpenShell across sandbox lifecycle operations
- Coding-agent setup: a starter prompt lets a local coding agent install NemoClaw step by step with approval before each command
