GhostFilter AI

Permissions & Privacy

What the SDK, CLI, browser extension, and connected sources can access.

npm SDK

By default, ghostfilter-ai has no external privileges. It does not automatically:

  • read files or directories
  • inspect shell history or environment variables
  • access email, browser tabs, or accounts
  • execute commands or tools
  • send scanned content over the network

It analyzes only the string passed by the host application. Remote network use occurs only when GHOSTFILTER_API_URL is configured and protect() runs an agent check.

CLI

The CLI reads only:

  • text supplied as a command argument
  • stdin supplied to pipe
  • the exact command string supplied to guard

guard analyzes a command. It never executes that command.

Browser extension

Ghosti Guard uses:

  • activeTab to inspect the current tab after a user action
  • scripting to extract selected or visible page text
  • local storage for API configuration
  • permission to the configured GhostFilter API origin

It has no background crawler and does not continuously record browsing.

Connected sources

Google, GitHub, and Slack connectors request read-only access for scanning. OAuth tokens are encrypted before storage. Public Convex queries strip token fields before returning connection metadata to the browser.

The Next.js OAuth callback and Convex scanner must use the same OAUTH_TOKEN_ENCRYPTION_KEY. Changing that key requires users to reconnect the affected provider. Outlook is prepared in code but is outside the judged production path.

Authentication

Web accounts are stored in Convex. Emails are normalized and unique. Passwords use per-user random salts and scrypt hashes; plaintext passwords are not stored.

Results are safety guidance, not a guarantee. Consequential actions should retain independent verification and human approval.

On this page