Agent onboarding adds responsibilities beyond the sandbox
NVIDIA NemoClaw supports OpenClaw, Hermes and LangChain Deep Agents Code inside OpenShell sandboxes. If you are deciding how to run one of those agents with controlled access, that integration is the reason to consider it. A generic ability to execute a process in a container would answer a smaller question.
The README describes guided onboarding, inference management, network policy and state persistence through its CLI. These components belong together operationally: an agent needs a model route and access to useful services, and you need to know what survives a restart or rebuild. Our lab record does not demonstrate those runtime behaviors. It records repository setup and a test command that never reached completion.
NemoClaw is explicitly alpha. The README says maintainers handle issues, discussions and pull requests on a best-effort basis without guaranteed response timelines. That is a concrete adoption limit for a team whose service depends on quick upstream support. Treat the supported-agent list as a starting point for an evaluation, rather than evidence of a service-level commitment.
What happened when we ran it
Our npm install succeeded in 65 seconds at commit 4160e2f on October 13, 2026, with 417 dependencies recorded and 447 MB on disk. The harness found no build script or target and skipped that step. Tests timed out after 900 seconds, so there is no completed-suite result to report.
The last lines show passing test files for plan validation, sandbox lifecycle and onboarding admission, among others. They also contain Node experimental SQLite warnings. A warning beside passing output does not identify why the overall command ran out of time. The supplied tail gives neither a final failing assertion nor a timeout diagnosis. We cannot convert it into a pass, or conclude that SQLite caused the timeout.
Our npm audit reported 18 known vulnerabilities: 2 critical, 7 high, 8 moderate and 1 low. The supplied record does not identify packages, exploit paths or whether affected code is reachable in a deployed sandbox. Those advisories deserve triage before adoption. They are not proof that a particular agent can escape its sandbox, and they do not constitute an isolation assessment.
The configured sandbox had 3 CPUs, 8 GB RAM and image lab-node:22. The label alone establishes neither the actual Node version nor the operating-system version. Current prerequisites specify a minimum of 4 vCPU. This older repository run therefore cannot verify operation on a currently supported host, and it cannot establish why the tests timed out.
Contributor setup does not create a runtime deployment
The README explicitly separates repository preparation from end-user installation. Its contributor instructions describe repository-local changes by default and separate options for exposing a development CLI or doing runtime validation. That distinction prevents a misleading conclusion from our successful dependency install: downloading packages is only part of getting an agent running.
The prerequisites document container-runtime choices with platform-specific qualifications. Docker is the primary onboarding path, while rootless Podman has its own supported configurations and limitations. Check the row for your actual host before following a generic quickstart. Our record does not evaluate either runtime provider.
Inference is another choice to settle before installation. A hosted provider needs its credentials and a permitted network route; a local inference service has separate model and memory requirements. The prerequisites distinguish these from the base host requirements. Neither the successful npm step nor the checkout's size tells you what a selected local model will require.
A policy decision belongs in the acceptance check
NemoClaw advertises network policy as part of the managed agent workflow. That makes a useful first runtime exercise quite specific: give the agent one required destination, then check how it handles a destination you have not approved. We have not performed that exercise. It would test the behavior you actually need from the selected agent and policy configuration.
Do the same for state persistence. Ask which agent data should survive a rebuild, then confirm that behavior on disposable material. The README's state-persistence claim supplies a feature to evaluate, not evidence that our timed-out repository checks covered recovery. A deployment trial should leave you able to explain both the permitted actions and the retained data.
The installer documentation also discusses Docker access and warns that membership in the Docker group has root-level impact. Decide which local accounts may operate that runtime. Sandboxing the agent does not remove the operator's responsibility for the host running the management tools.
Direct runtime control may fit a smaller job
NVIDIA OpenShell is the underlying policy runtime. Its README says the default sandbox image has no agent installed. Choose that direct path when your team wants to supply its own agent setup and work with the runtime controls itself. NemoClaw becomes interesting when its supported-agent onboarding saves integration work you would otherwise own. We have no comparative operating-effort measurement.
Anthropic Sandbox Runtime restricts filesystem and network access for arbitrary processes without requiring a container. That is a candidate for restricting a local command or process. Its README labels it a research preview with evolving APIs and configuration, so it is not an answer to NemoClaw's alpha status if guaranteed stability is your requirement.
Choose one supported agent and a host listed in the current platform guidance. Before connecting useful credentials, verify a real inference request, an intentional policy denial and the state you expect to recover after rebuilding.

