Skip to content

Privacy

Jump to bottom
Chuyue Wang edited this page Mar 24, 2026 · 2 revisions

Privacy

What Cortex Collects and Where It Goes

Cortex is designed to keep biometric data on your machine. Here is an exact account of every piece of data the system handles.

Webcam / Biometrics

Data What happens
Video frames Processed in memory at 30 FPS. Never written to disk. Never sent anywhere.
Face images Used to extract landmarks and RGB traces. Immediately discarded after processing.
Heart rate (BPM) Computed locally. Stored in memory for state classification. Never sent to the LLM or any external service.
HRV (RMSSD) Same as heart rate — local only.
Blink rate Local only.
Head pose / posture Local only.

No biometric data ever leaves your machine.

Behavioral Telemetry

Data What happens
Mouse movements Aggregated to velocity/jitter statistics at 1 Hz. Raw events are not stored.
Keyboard activity Aggregated to burst scores and interval variance. Keystroke content is never captured — only timing patterns.
Window/tab switches Switch rate is tracked (count per minute). Titles are read for context only when an intervention is triggered.

What the LLM Sees

When Cortex triggers an intervention, it sends a context packet to the LLM. This contains only workspace context:

  • Current cognitive state label and confidence (e.g., HYPER 0.91)
  • Active file path and visible code range from VS Code
  • Active diagnostics (error messages, line numbers) from VS Code
  • Tab titles and URLs from Chrome (sampled, max 30)
  • Recent terminal output (text only, not command history)

The LLM never receives:

  • Heart rate, HRV, or any physiological measurement
  • Video frames or face images
  • Mouse/keyboard content
  • Clipboard contents

Data Storage

Storage What Location
Baselines Calibrated resting HR, HRV, blink rate, posture storage/baselines/ — local only
Session features Aggregated feature vectors (7-day retention) storage/features/ — local only
Activity tracker Learning progress (platform, URL, position) storage/activity/ — local only
Bandit weights Learned intervention preferences storage/eval/ — local only
Error logs Crash reports and exceptions storage/errors/ — local only, 90-day retention

Redis is used if available, with automatic in-memory fallback. No cloud sync.

API Key Security

The Azure OpenAI API key can be stored in .env (development) or in macOS Keychain (recommended for production):

security add-generic-password -s cortex.azure_openai -a default -w YOUR_KEY

Then in .env:

CORTEX_LLM__AZURE__USE_KEYCHAIN=true

The key is never logged, never included in WebSocket messages, and never sent to the extension.

Browser Extension Permissions

The extension requests broad permissions to function. Here is exactly why each is needed:

Permission Why
activeTab Read current tab URL for context building
scripting Inject the intervention overlay
tabs Read all tab titles/URLs for the LLM context packet
tabGroups Create named tab groups (intervention action)
storage Save quiet mode state, bandit feedback, activity
alarms Schedule focus session end notifications
bookmarks Bookmark-and-close action
webNavigation Detect navigation for the activity tracker
nativeMessaging Start/stop the Python daemon
<all_urls> Inject the intervention overlay on any page

Consent-Gated Autonomy

Cortex cannot take any autonomous action without earned trust. The consent ladder has 5 levels per action type:

OBSERVE → SUGGEST → PREVIEW → REVERSIBLE_ACT → AUTONOMOUS_ACT
  • Cortex starts at SUGGEST — it shows recommendations but takes no action
  • 5 user approvals of the same action type escalate to the next level
  • 3 user rejections de-escalate by one level
  • No tab is ever closed, no code is ever folded, without the user either approving in real-time or having previously granted that action type sufficient trust

All interventions are reversible. Workspace state is snapshotted before any action and can be restored with one click.

Clone this wiki locally