Security

The Claude API key

• Stored only on the server, encrypted with AES-256-GCM using a key derived from APP_SECRET.
• Never returned by any API; at most a masked hint like sk-ant-…a1b2 is shown.
• Never logged. AI errors shown to users are rewritten so request details cannot leak.
• The browser never talks to Anthropic directly; every AI request goes through the Go backend.

Accounts

• Passwords are hashed with bcrypt.
• Sessions use random tokens in httpOnly, SameSite=Lax cookies (Secure over HTTPS); only a hash is stored.
• OAuth uses state and PKCE, and links accounts only by verified email.
• Sign-in endpoints are rate limited.
• State-changing requests from other sites are rejected.

User-generated sites

Previews and published sites are served with a Content-Security-Policy sandbox. They run in an isolated origin, so their scripts cannot read the studio's cookies or call its API as the signed-in user. Preview links are signed and expire.

Limits

Per-file and per-project size limits, request size limits, a per-user request rate limit, one AI run per project at a time, and an optional daily token limit per user.