AI, permissions & data controls

1. Ari’s architecture and the public preview

Ari is a configurable AI companion with voice, text, an interactive glass orb, a local context library and reviewed capability modules. Its Mac Core coordinates conversations and tool workflows. The iPhone companion privately connects to the same running Mac rather than independently duplicating every model and tool.

The website demonstrates that design through locally scripted scenes. It is not running the native Ari against your device and is not a public API for research, email or app control. Particular releases and launch plans must identify exactly which capabilities they include.

2. Model choice and provider behaviour

Configured integrations include OpenAI, xAI/Grok, Z.AI and compatible installed local models through LM Studio and Bionic. Vision, tool use, response quality, hardware needs and available context depend on the chosen model and runtime.

A local conversation model can reduce model-provider transfers, but a live web search, online account, optional sync or cloud browser still has its own network path. Review the provider agreement, account settings and usage costs. A subscription to an Ari module does not automatically include every external provider account or fee.

3. Tool authority and visible outcomes

Tools are bounded by their implementation, workspace scope, service credentials and operating-system permissions. Modules and MCP connections expose particular capabilities; they are not general authority to operate every application. The shipped GitHub and Supabase connections reviewed for this site are read-only.

For email, inspect the displayed recipients, subject and full draft before the later send confirmation. For a coding task, review the scoped changes and actual build/test result. For browser work, taking control can pause the automation. Use the stop or revoke controls relevant to the workflow.

The orb’s colour, movement and speech communicate presence and activity; they are not proof that a tool succeeded. Check task status and the returned artifact, citation, file or provider response. A meeting’s proposed action item remains a proposal until an action actually completes.

4. Memory, documents and reusable procedures

Ari can retrieve saved conversations, research, meeting notes and documents from a local knowledge library. User-confirmed personal facts and reusable skill cards can help a later task. Retrieval is selected context, not perfect memory or continuous training on everything shown on the device.

Optional Cloud Memory can sync selected collections and retain a searchable local mirror. Review the collections and destination before enabling it. A procedure stored in a skill card does not grant a new permission, override a service’s access rules or authorise an unrelated action.

5. Voice, meeting and screen controls

Voice-profile setup helps identify the speaker. Optional Hearing training is separate, off by default, and has its own local data and removal controls. Voice-profile or audio information may be sensitive; consent, security and retention need to match the actual deployment and law.

Start meeting recording/transcription with the necessary participant notice and consent. iPhone screen sharing uses Apple’s explicit picker and current device/OS requirements. A started background phone voice session remains subject to iOS and ends when the app is force-quit. Do not treat a Live Activity as a promise of indefinite background execution.

6. Browser and optional remote sessions

Ari’s native Mac browser has its own session, tabs and page assistance. It does not silently inherit all Chrome or Safari account cookies. Give explicit instructions about the browsing context and keep authentication or material account decisions under your control.

An optional cloud browser is for public, read-only work with a watch-only view. It is not an unrestricted remote desktop. A paid session needs approval; hiding the viewer does not end it. Use Stop and check the actual session status. Keep the host Mac awake with Ari running.

7. Accuracy, safety and human judgement

Review factual claims against current sources and avoid sole reliance for consequential decisions. A model may misunderstand a screen, miss context, follow a misleading source or describe a step that did not run. Use a safe test dataset and verify a task before trusting it with important work.

Tell us about an unclear capability or unexpected result at info@nightwingdigital.co. The report should identify the version, task and observed outcome without exposing secrets. We will distinguish an implementation defect, a missing permission, provider behaviour and a model error rather than imply a single animation proves reliability.