Autonomous AI agents that execute terminal commands or run code loops need strict safety boundaries before running shell operations.

We built Veltra, an open-source pre-execution security engine written in Python and FastAPI.

🛡️ What Veltra Does:

  • Sub-3ms Local Inspection: Inspects shell commands and environment variable access before execution.
  • Destructive Pattern Blocking: Catches dangerous commands like rm -rf, unauthorized env overrides, and malicious payload injections.
  • FastAPI Scan Service: Exposes a clean /api/v1/scan endpoint returning ALLOWED or SECURITY_VIOLATION verdicts.

🧪 Real-World Integration

We recently verified Veltra against browser-use terminal execution workflows, passing all 48 test suites cleanly with median library overhead between 1.3ms and 3.3ms.

Would love to hear any thoughts or feedback from the developer community!****