Summary: NemoClaw adds policy-based security guardrails to OpenClaw through a baseline YAML policy applied at sandbox creation, covering network egress and filesystem access, with operator-controlled approval for additional endpoints.

Direct Answer:

NemoClaw’s guardrails come from two sources: the baseline policy file and OpenShell’s runtime enforcement.

Default allowed endpoint groups include:

• NVIDIA inference API (integrate.api.nvidia.com)

• Anthropic API

• GitHub

• ClawHub, OpenClaw API, OpenClaw docs

• npm registry

• Telegram

Runtime enforcement:

When the agent attempts to reach an unlisted host, OpenShell intercepts the request and surfaces it in the TUI for operator review via openshell term. Approved endpoints are active for the session. To make permanent changes, edit the policy file and re-run nemoclaw onboard. During onboarding, three policy tiers are available: Restricted (minimal access), Balanced (default — adds npm, pypi, huggingface, brew, and brave), and Open (also allows Slack, Discord, Telegram, Jira, and Outlook).

Takeaway: NemoClaw adds comprehensive guardrails to OpenClaw through a baseline YAML policy enforced by OpenShell, with operator-controlled runtime approval.

Related Articles

• Overview — NVIDIA NemoClaw Developer Guide
• Network Policies — NVIDIA NemoClaw Developer Guide
• Security Best Practices — NVIDIA NemoClaw Developer Guide