Deutsch
NemoClaw provides a way to run AI agents with defined access to files, networks and model services. It puts the execution environment around the agent under explicit controls, which is useful when the agent can do more than answer questions.
A team evaluating an assistant might allow it to read one project directory and contact one model endpoint, while blocking unrelated files and services. The important test is whether the configured environment permits the intended task and rejects the access it should not have.
What NemoClaw is
NemoClaw is NVIDIA's open-source reference stack for running supported agents inside NVIDIA OpenShell sandboxes. The NVIDIA repository describes guided setup, inference routing, network policies, managed integrations, snapshots and lifecycle operations.
OpenClaw is one supported agent. NemoClaw supplies the surrounding setup and management, while OpenShell provides the sandbox environment. The agent still needs a model, instructions and tools to perform its task.
Use the upstream documentation to check the supported agents and platforms for the release you plan to run. The NemoClaw self-hosting guide provides the VPS setup path.
What the controls are for
The runtime determines which resources the agent can reach. Review the effective configuration across several areas:
| Area | Decision to make | Example check |
|---|---|---|
| Filesystem | Which directories can be read or written | An unrelated host directory is inaccessible |
| Network | Which destinations may be contacted | An unapproved endpoint is blocked |
| Inference | Which model service receives requests | Requests use the intended provider |
| Integrations | Which tools and credentials are available | A read-only task cannot update records |
| Lifecycle | How the environment starts, stops and recovers | A restart preserves the required state |
| Logs | What activity is recorded and retained | A denied action can be investigated |
These are checks to perform on your chosen configuration, rather than a promise that every installation has the same controls. Read the official architecture and security documentation before depending on a particular boundary.
A sandbox also has limits. Access granted to an external tool can still be consequential, and an allowed file or response may contain misleading instructions. Restrict permissions and review outputs even when the execution environment is isolated.
How a controlled task can run
Start by defining the task and the minimum access it needs. Prepare the agent environment, model connection and policy, then run a harmless example with representative inputs.
Check both successful and denied actions. If a documentation task can read its project files but cannot write outside the workspace or contact unrelated hosts, retain those results with the policy version used for the test.
When a legitimate task needs more access, review the specific destination or directory. Broadly disabling a restriction to make one task work can remove the protection the environment was meant to provide.
Practical use cases
Testing an agent before giving it broader access
Use disposable files and test credentials to evaluate tool use. Confirm what the agent can read, write and contact before connecting it to valuable data or shared systems.
Comparing permission policies
A team can run the same workflow under different approved policies and record which permissions it actually needs. This is useful when deciding whether a proposed integration requires read access, write access or a separate approval step.
Evaluating OpenClaw workflows
OpenClaw supplies the assistant behaviour and integrations. Running it through NemoClaw lets a team evaluate those workflows with OpenShell sandbox controls and a repeatable configuration. Test the integration itself as well as the enclosing environment.
Reproducing and investigating a run
Keep the agent version, policy, relevant configuration and logs together. They help explain which access was available when a task ran and make it easier to reproduce an unexpected result.
Logs alone may not capture every external effect. For a tool that changes another system, also check that system's resulting state and audit records.
How NemoClaw differs from OpenClaw
OpenClaw is the assistant that receives requests and uses tools. NemoClaw helps set up and operate supported agents inside a controlled OpenShell environment. They occupy different parts of the stack rather than being interchangeable chat interfaces.
Choose the additional runtime controls when they address a requirement you can test. A simple local experiment may not need the same setup as an agent connected to team repositories or business systems.
When a VPS is suitable
A VPS can keep the runtime, policy and logs in a dedicated environment with remote access. It can also make an experiment easier to rebuild than a collection of changes on a personal workstation.
Check the upstream prerequisites for your exact release. Operating-system support, container capabilities, memory, storage and inference choices all affect suitability. Do not assume that any low-cost VPS can run the full stack or a local model.
Keep the host patched, restrict SSH, separate administrative accounts and protect credentials outside version-controlled policy files. Configure restart behaviour and monitor both the host and the sandboxed workload. Retain independent copies of the state you need to restore.
The VPS guide covers setup; validate its steps against the current upstream requirements before applying them to an existing server.
Plan failure and recovery
Decide who can approve policy changes and how to revoke a tool credential. Set log retention and access rules, especially if prompts or tool outputs contain confidential information.
Test how to stop a misbehaving task, disable an integration and return to a known policy. A snapshot can help with some local changes, but it cannot undo a message already sent or a remote API update. Keep human approval for sensitive external actions and document their recovery options separately.
On a self-managed VPS, your team operates the host, agent software, policies and integrations. A repeatable setup makes that work easier to inspect; it does not remove the work.
Frequently asked questions
What is NemoClaw used for?
It helps set up and manage supported AI agents in OpenShell sandboxes, with controls such as network policies and inference routing. It is useful when those access boundaries matter to the workflow.
Is it the same as OpenClaw?
No. OpenClaw is an agent that can run within the stack. NemoClaw manages the surrounding setup and lifecycle, using OpenShell for sandboxing.
Does it make every agent action safe?
No. The result depends on the effective policy, granted credentials, external tools and the task itself. Test restrictions and retain review for consequential actions.
Why use a VPS?
A compatible VPS can provide persistent state and remote access in a dedicated environment. Compare the actual runtime requirements and operational work before selecting a plan.
