Security & Trust Center
Architecture Overview
SorvoAI uses a client-server architecture with a native desktop application (Electron) and a centralized API server. There is no web application by design — native apps provide stronger security boundaries, system-level integration, and reduced attack surface compared to browser-based alternatives.
Client (Your Device)
- Native desktop app (Windows, macOS)
- Screen analysis runs locally — raw screenshots never leave your device
- Authentication tokens stored in OS keychain via Electron safeStorage
- Auto-update via signed NSIS installers with SHA-512 verification
Server (Cloud)
- API server on dedicated VPS (not shared hosting)
- PostgreSQL database with pgvector for semantic search
- Redis for job queues, caching, and real-time pub/sub
- BullMQ for async job processing with dead-letter queue
Data Flow: What Stays Local vs. What Leaves
Stays on your device
- Raw screen captures and screenshots
- Authentication tokens and session keys
- Local application cache
- Offline data cache (encrypted)
Transmitted to server (encrypted via TLS 1.3)
- Chat messages and responses
- Screen analysis metadata (app name, window title — not pixels)
- Calendar events, email metadata, and synced data
- Memory artifacts, entity graph, and intelligence outputs
- File uploads (for knowledge base)
Encryption Model
- In transit: All API traffic encrypted via TLS 1.3
- At rest: Database encrypted at the volume level
- Enterprise E2E: Optional client-side end-to-end encryption via libsodium (X25519 + XChaCha20-Poly1305)
- Token storage: OS keychain integration (Windows DPAPI / macOS Keychain)
- Sensitive fields: API keys, OAuth tokens, and SMTP credentials encrypted via AES-256-GCM before storage
Authentication & Access Control
- JWT-based authentication with short-lived access tokens (15 min) + long-lived refresh tokens
- Brute-force protection with progressive lockout
- Enterprise: SAML 2.0, OIDC, Okta, Azure AD support
- Enterprise: IP allowlisting for network-level restriction
- Role-based access control for team workspaces
Data Retention & Deletion
- Events archived after 90 days (soft delete, recoverable)
- Working memory expires after 24 hours
- Screen events retained per user preference (default: 30 days)
- Full account deletion with 72-hour grace period
- GDPR data export available (ZIP format, all user data)
AI Model Usage
- Primary model: Qwen 3 235B Instruct (self-hosted inference)
- Reasoning: Qwen 3 Thinking
- Enterprise: Claude Sonnet 4 (via Anthropic API)
- We never train models on your data. Your conversations, memories, and files are never used for model training.
- Content boundaries prevent sensitive data from reaching LLM contexts
Subprocessors
- Hostinger VPS: Application hosting (dedicated, not shared)
- OpenAI API: Embedding generation (text-embedding-3-small)
- SendGrid: Transactional email (agent email, webhooks)
- PostHog: Product analytics (anonymized usage data only)
- Sentry: Error monitoring (no PII in error reports)
- Google APIs: OAuth, Calendar sync, Gmail sync (user-authorized)
Compliance Roadmap
- GDPR: Data export, deletion, and consent management implemented
- SOC 2 Type II: On roadmap for 2026 Q4
- HIPAA: Not currently supported (healthcare use not recommended)
- ISO 27001: Under evaluation
Reporting Vulnerabilities
If you discover a security vulnerability, please report it to security@sorvoai.com. We take all reports seriously and will respond within 48 hours.