Raw biological sequences are local-only. The server and MongoDB may receive identity, preferences, project names, document names, lengths, alphabet/topology, and opaque local storage keys. They must not receive sequence strings, imported file bodies, derived constructs, clipboard contents, or analysis results capable of reconstructing a sequence.

This is enforced with strict allow-list schemas and small request limits. There is no generic document upload endpoint.

Production auth uses Better Auth with HTTP-only cookies and MongoDB. BETTER_AUTH_SECRET must be at least 32 high-entropy characters. Google OAuth is enabled only when both Google variables are configured. Guest mode is explicit and never simulates a production session.

Never commit .env, OAuth secrets, Atlas credentials, cookies, or exported user data.

• Read-only and UI-mutating tools declare their effects truthfully.
• Imported/user-authored output is marked with untrustedContentHint.
• Inputs and outputs are bounded.
• WebMCP calls use the same validated application commands as the visible UI.
• Annotation and cloning changes are staged for human approval.
• Tool descriptions are developer-authored and never interpolate imported content.

Until a dedicated security mailbox exists, open a private GitHub security advisory on the repository. Do not include private sequence data in a report.