SafeClaw

An observability and control layer for autonomous AI agents that monitors shell commands, file access and network calls, classifies threats in real time and blocks dangerous, exfiltrative or unauthorized actions before execution. It is for developers and IT teams in businesses that run autonomous agents and need to enforce safety controls. It is delivered as a SaaS offering installed via npm with real-time interception, human-in-the-loop approvals and audit logging.

Key features

• Monitor shell commands in real time
• Block dangerous inputs
• Sandbox file system
• Real-time threat classification
• Secret and credential detection
• Data exfiltration blocking
• Prompt injection detection
• Destructive operation blocking
• Privilege escalation detection
• Supply chain verification
• Sensitive file access control
• Network activity monitoring
• MCP poisoning detection
• 12-layer security posture scoring
• Human-in-the-loop approvals
• Full audit trail logging
• Skill and MCP security scanning
• Per-agent access control
• Per-directory filesystem controls
• Custom restricted command patterns