Our philosophy#
Ellie is designed to assist, not override. She accelerates tasks users are already authorized to perform, while enforcing strict security, consent, and visibility boundaries.
Ellie does not act independently, does not escalate privileges, and does not access sensitive secrets.
Identity & access control#
Operates under your identity
Ellie operates exclusively using the same JWT and permissions as the signed-in user.
- No service accounts
- No elevated privileges
- No cross-account access
If a user cannot perform an action manually, Ellie cannot perform it either.
Role-aware by design
Ellie fully respects Hostney’s existing permission model:
- Account-level access
- Domain-level access
- Read-only vs write access
- MSP delegate boundaries
No additional AI-specific roles are required.
Explicit user consent#
Destructive actions require opt-in
Potentially destructive actions (such as deletes, rebuilds, or irreversible changes) require explicit user-enabled consent in settings. Ellie will never perform destructive operations unless this consent is enabled.
Public IP usage is disclosed
Any operation that requires disclosing or using a public IP address is clearly disclosed and must be explicitly allowed by the user. Consent can be revoked at any time.
No access to sensitive secrets#
Ellie cannot access:
- Files or file contents
- Passwords or credentials
- SSH keys
- Private SSL keys or certificates
All sensitive material remains inaccessible to AI processes.
Guided, not blind automation#
Ellie does not perform hidden background actions. Every task is executed using a visual, step-by-step execution plan, which may include:
- Page navigation
- Account and domain selection
- Form interactions
- User confirmations when required
Users always remain aware of what is happening.
Operating modes#
Ellie supports two execution preferences:
Visual mode (default) Ellie performs actions by navigating through the actual interface, clicking buttons, filling forms, and showing you exactly what’s happening. You can follow along and see each step.
Execute mode (opt-in) For faster execution, you can enable “Execute in chat when available” in settings. Ellie will call APIs directly to complete tasks without navigating the UI. Ideal for repetitive or bulk operations.
Tutorial mode For specific pages, Ellie can walk you through the interface step-by-step, highlighting elements and explaining each part without making any changes. Ask “Show me how this page works” to start a tutorial.
Full audit logging#
Every Ellie-initiated action is logged with:
- User identity
- Parsed intent
- Execution steps
- Active consent settings
- Timestamp and target resources
Audit logs provide a complete, reviewable record of AI activity.
Deterministic & constrained behavior#
Ellie does not improvise actions. All operations are executed using:
- Whitelisted actions
- Validated execution maps
- Strict UI-level constraints
Ellie cannot perform actions outside her defined capabilities.
Designed for compliance#
Ellie’s architecture aligns with modern security and compliance expectations, including principles found in:
- SOC 2
- CMMC
- ISO-aligned access control models
Ellie is built to be explainable, auditable, and predictable.
Transparency first#
Ellie will always explain:
- What she intends to do
- Why it’s being done
- What permissions are required
Users remain in control at all times.
In short#
- Ellie never exceeds user permissions
- Ellie never accesses secrets
- Ellie never acts without consent
- Ellie never hides actions
- Ellie always leaves an audit trail
That’s how AI should work.