Architecture Details

NemoClaw combines a host CLI, an in-sandbox integration layer, and a versioned YAML blueprint that defines the sandbox image, policies, and inference profiles applied through OpenShell.

System Overview

NVIDIA OpenShell is a general-purpose agent runtime. It provides sandbox containers, a credential-storing gateway, inference proxying, and policy enforcement, but has no opinions about what runs inside. NemoClaw is an opinionated reference stack built on OpenShell that handles what goes in the sandbox, prepares agent-specific integration, and makes the setup accessible.

Deployment Topology

The logical diagram above shows how components relate. This section shows what actually runs where on the host. NemoClaw’s default Docker-driver topology does not place the sandbox in an embedded k3s cluster. On Linux and Apple Silicon macOS, NemoClaw starts the OpenShell Docker-driver gateway and creates the sandbox as a Docker container. The gateway normally runs as a host process; Linux hosts that need the gateway compatibility patch may run the same gateway binary inside a small container. In both Docker-driver modes, the sandbox is a Docker container, not a Kubernetes pod. Legacy non-Docker-driver installs still use the k3s-based gateway path; the diagram below shows the standard Docker-driver topology.

Layering from top to bottom:

NemoClaw never gives the sandbox a raw provider key. At onboard time it registers credentials with OpenShell’s provider/placeholder system, and the L7 proxy substitutes the real value into outbound requests at egress. The CLI helper isInferenceRouteReady (in src/lib/onboard.ts) is a host-side readiness check used by the resume flow to decide whether the active route already covers the chosen provider and model. It is not a runtime component.

For the DGX Spark-specific variant of this topology (cgroup v2, aarch64, unified memory), refer to the NVIDIA Spark playbook.

NemoClaw Agent Integration

NemoClaw integrates with each supported agent through a runtime layer that adapts the agent to OpenShell-managed providers, policies, and sandbox state. The concrete files differ by agent because each runtime has its own plugin system, config format, state layout, and startup command.

The OpenClaw integration is a thin TypeScript plugin that runs in-process with the OpenClaw gateway inside the sandbox. Before an OpenClaw turn starts, the plugin prepends a short context block with the active sandbox name, sandbox phase, network policy summary, and filesystem policy summary. When the policy or phase changes during a session, the plugin sends a smaller update block instead of repeating the full context.

The Hermes integration follows the generic agent-manifest path instead of the OpenClaw plugin package path. The manifest declares Hermes’ binary, health probe, config directory, state directories, messaging support, and OpenAI-compatible API endpoint. The build-time config generator turns NemoClaw onboarding choices into Hermes YAML and environment files, and the Hermes plugin manifest exposes NemoClaw tools and an on_session_start hook.

NemoClaw Blueprint

The blueprint is a versioned YAML package with its own release stream. The runner resolves, verifies, and applies the blueprint through the OpenShell CLI. The blueprint defines the sandbox shape, default policies, and inference profiles; the runner performs the OpenShell operations.

nemoclaw-blueprint/
├── blueprint.yaml                  Manifest: version, profiles, compatibility
├── model-specific-setup/           Agent-scoped model/provider compatibility manifests
├── router/                         Model Router config and routing engine
├── policies/
│   └── openclaw-sandbox.yaml       Default network + filesystem policy for the OpenClaw profile

Hermes keeps its agent-owned image, plugin, config, entrypoint, and policy additions under agents/hermes/. The default Hermes policy starts from agents/hermes/policy-additions.yaml, and the permissive variant used by shields down lives in agents/hermes/policy-permissive.yaml.

The current blueprint runner implementation lives in the nemoclaw/ TypeScript package:

nemoclaw/src/blueprint/
├── runner.ts                       CLI runner: plan / apply / status / rollback
├── ssrf.ts                         SSRF endpoint validation (IP + DNS checks)
├── snapshot.ts                     Migration snapshot / restore lifecycle
├── state.ts                        Persistent run state management

Blueprint Lifecycle

1. Resolve. The integration layer locates the blueprint artifact and checks the version against the OpenShell and agent runtime constraints in blueprint.yaml.
2. Verify. The integration layer checks the artifact digest against the expected value.
3. Plan. The runner determines what OpenShell resources to create or update, such as the gateway, providers, sandbox, inference route, and policy.
4. Apply. The runner executes the plan by calling openshell CLI commands.
5. Status. The runner reports current state.

Sandbox Environment

Normal NemoClaw onboarding builds from the ghcr.io/nvidia/nemoclaw/sandbox-base base image and layers the NemoClaw runtime Dockerfile on top. The direct blueprint runner still carries a pinned OpenShell Community OpenClaw image for legacy openshell sandbox create --from compatibility. Inside the sandbox:

• The selected compatible agent runs with the NemoClaw integration layer installed or generated for that agent.
• Inference calls are routed through OpenShell to the configured provider.
• Network egress is restricted by the baseline policy for the selected agent profile.
• Filesystem access is confined to /sandbox and /tmp for read-write access, with system paths read-only.
• NemoClaw injects sandbox and policy context into agent turns when the selected agent supports runtime context hooks, so the agent can report policy blocks accurately.
• The image exposes a Docker health check that probes the in-sandbox gateway, so container runtimes can report whether the agent service is responding.
• The image includes common runtime compatibility helpers such as Homebrew and a python to python3 symlink for tools that still invoke python.

Inference Routing

Inference requests from the agent never leave the sandbox directly. OpenShell intercepts them and routes to the configured provider:

Compatible agent (sandbox)  ──▶  OpenShell gateway  ──▶  Provider endpoint

When you select the Model Router provider, the OpenShell gateway routes to a host-side router process instead of a single upstream model. The router selects from the configured pool, then calls the upstream NVIDIA endpoint with the credential held outside the sandbox.

Some model and provider combinations need agent-specific compatibility setup. NemoClaw keeps those declarations under nemoclaw-blueprint/model-specific-setup/<agent>/ so fixes for each supported agent can be tested and reviewed independently.

Refer to Inference Options for provider configuration details.

Provider Credential Storage

Provider credentials live in the OpenShell gateway store, not on the host filesystem. NemoClaw never writes them to host disk; the OpenShell L7 proxy injects values at egress. See Credential Storage for the inspection, rotation, and migration flow.

Host-Side State and Config

NemoClaw keeps non-secret operator-facing state on the host rather than inside the sandbox.

The following environment variables configure optional services and local access.

For normal setup and reconfiguration, prefer nemoclaw onboard over editing these files by hand. Do not treat NEMOCLAW_DISABLE_DEVICE_AUTH as a runtime setting for an already-created sandbox.