When you give an AI agent control of your mouse and keyboard, the obvious question is: what does it see, and where does that data go? At Oprel, we built the answer into the core architecture from day one, not as an afterthought.
The Privacy Sieve
Every screenshot Oprel captures before sending anything to the AI is first processed locally by the Privacy Sieve. This is a local-first redaction layer that runs entirely on your device. It uses 30+ regex patterns combined with Tesseract OCR to scan every frame for personally identifiable information before any data leaves your machine.
What gets masked:
- Email addresses
- Phone numbers
- Credit card and bank numbers
- Passwords and API keys
- National ID numbers, passport numbers
- Names in form fields
- Any field labelled "confidential", "private", or "secret"
The Privacy Sieve runs before the screenshot is encoded and transmitted. The AI model never sees unmasked PII, ever.
Local-First Architecture
Oprel does not run in the cloud. The agent runs as a native desktop application. Screenshots are taken locally, processed locally, and only the redacted, sieved version is sent to the AI model for reasoning. No raw screen data is stored on Oprel's servers.
The Three-Layer Safety System
Beyond privacy, Oprel has a three-layer system that governs what actions it can take:
Layer 1: The Governor
A command blacklist that blocks specific high-risk operations before they are even considered by the AI. The Governor operates at the intent-classification level.
Layer 2: The Safety Guard
Every planned action is classified into one of four risk levels: Safe, Ambiguous, Risky, or Blocked. Ambiguous and Risky actions trigger a Human-in-the-Loop pause where you approve or deny before execution continues.
Layer 3: Hard Blocks
Certain actions are permanently, unconditionally blocked regardless of any instruction: formatting drives, deleting system directories, disabling security software. These cannot be unlocked, overridden, or prompted around.
Human-in-the-Loop
Oprel pauses execution the moment it detects user input, keyboard or mouse, during a task. You are always in control. You can deny any action, cancel any task, and intervene at any point. The agent is designed to assist, not to operate autonomously beyond your awareness.
Privacy and safety are not features we added. They are the foundation the product is built on. If you have questions about how Oprel handles specific data types or use cases, reach out to support@getoprel.com.