Security & compliance

Implemented safeguards

• Separate operator and agent credentials; hashed, individually revocable scoped agent keys.
• Server-owned tool registry and model allowlists with strict supported inputs.
• Human approval for sensitive registered tool actions, expiry and replay protection.
• Transactional budget/token reservations, idempotent requests and persistent emergency stop.
• Same-origin checks, bounded request bodies, request limits and storage caps.
• Safe DOM text rendering and restrictive browser security headers.
• Model content is not retained in model history; metadata and supported tool arguments/results are retained.
• Presale consent records, no-card signup, deletion controls and clear beta disclosures.

Boundaries that matter

There is one private workspace, not isolated customer organizations. Access controls apply only to routed calls. In-flight actions cannot be recalled, external tools are not controlled, and AI-generated text still needs human review. Spending reservations rely on checked reference rates and are not provider invoice guarantees. No autonomous tool-execution loop is enabled.

Before regulated or sensitive use

The current beta has no customer DPA, per-tenant isolation, SSO/MFA, application-level encrypted content storage, independently verified security assessment, or automated full retention program. Do not use it for sensitive or regulated data. Assess provider terms and your own obligations before connecting business workflows. Public legal pages describe current practices; they are not certification evidence.

Report a concern

Contact mvmtas@gmail.com privately. Do not publish tokens, customer information or exploit details in public issue reports.