OpenClaw Was Dangerous… Until NemoClaw Stepped In (OpenSource)

Key Concepts

• Nemo Claw: An enterprise-grade security wrapper for Open Claw that provides sandboxing, network policy enforcement, and audit logging.
• Open Claw: An AI agent framework consisting of a gateway, reasoning engine (LLM), memory, skills, and scheduled tasks.
• OpenShell: The security dashboard and terminal interface used to monitor, log, and configure network policies for Nemo Claw.
• Sandbox: A restricted execution environment where the AI agent operates, preventing unauthorized file access and data exfiltration.
• Inference Router: A mechanism that routes API calls through a secure gateway to prevent sensitive credentials from entering the sandbox.
• Zero Trust/Denied by Default: A security posture where all network traffic and file access are blocked unless explicitly whitelisted via policy.

1. Overview and Purpose

Nemo Claw addresses critical security vulnerabilities in Open Claw, where malicious "skills" in AI marketplaces have been known to steal API keys and browser credentials. Nemo Claw acts as "Mission Control" for the Open Claw "astronaut," ensuring that no data leaves the environment without approval and that all actions are logged and monitored.

2. Technical Architecture

• The Three Components:
  1. Open Claw: The core agent framework (Gateway, Reasoning Engine, Memory, Skills, Scheduled Tasks).
  2. OpenShell: The security layer that monitors telemetry, logs decisions, and enforces flight rules.
  3. Nemo Claw: The orchestration layer that builds the mission, sets policies, and manages the sandbox.
• Security Model: Unlike standard Open Claw, which has full user account permissions, Nemo Claw restricts the agent to specific sandbox and temporary folders. It intercepts API keys at the gateway level so they never enter the sandbox, and it strips personal data from requests before they leave the environment.

3. Step-by-Step Implementation

1. Prerequisites: Ensure Docker, Git, and Curl are installed on a Linux-based machine (e.g., NVIDIA DGX Spark).
2. Environment Setup: Export the NVIDIA_API_KEY to the terminal.
3. Installation:
   • Install OpenShell and set environment variables.
   • Clone the Nemo Claw repository.
   • Make the installation scripts executable and run the setup.
4. Configuration: Nemo Claw automatically detects local hardware (e.g., NVIDIA GPU) and configures the inference.local routing.
5. Policy Management:
   • Access the OpenShell terminal (openshell term).
   • View logs (l) to identify blocked requests.
   • Edit the openclaw_sandbox.yaml file to add specific network policies (e.g., whitelisting a weather API URL).
   • Apply the policy using the command line to update the sandbox rules.

4. Key Arguments and Perspectives

• Security by Design: The presenter argues that standard AI agents are inherently insecure because they run with full system permissions. Nemo Claw mitigates this by enforcing a "denied by default" policy.
• Hardware Agnostic: While optimized for NVIDIA hardware, Nemo Claw is hardware-agnostic and can run local models (e.g., via Ollama) or external providers, ensuring privacy regardless of the underlying infrastructure.
• Auditability: Every "allow" or "deny" decision is logged in real-time, providing the transparency required for enterprise-grade deployments.

5. Notable Quotes

• "Think of it like a space mission where Open Claw is the astronaut, OpenShell is the spacecraft, and Nemo Claw is the mission control which builds the mission, sets flight rules, monitors every tele-signal, and logs all decisions."

6. Synthesis and Conclusion

Nemo Claw transforms Open Claw from a potentially vulnerable agent framework into a secure, enterprise-ready solution. By implementing a sandbox, intercepting API keys at the gateway, and requiring explicit network whitelisting, it effectively neutralizes the risk of malicious skills. The combination of real-time monitoring via OpenShell and strict OS-level network policies provides a robust framework for users to run AI agents privately and securely, whether using local models or external frontier models.