Cast by Stella private beta

Private AI for decisions.

A local-first decision workspace for context, evidence, assumptions, risk, and next actions. Built for users who want AI support with private reasoning kept on-device where the beta build supports it.

Decision workspacelocal first
Decision: respond now or wait for missing context.
Assessment: wait for the missing fact. The next action depends on confirmed context, not a guess.
FactsAssumptionsRiskNext action
On-device where supported User-controlled memory Decision structure Built for decisions

01 The workflow

Decision support with a reviewable structure.

The assistant captures context, asks for the missing fact that would change the recommendation, and separates facts, assumptions, unknowns, risk, and next action.

01Capture

Text, voice, screenshot, photo, file, deadline, or other evidence the user chooses to provide.

02Frame

Classify the problem type, stakes, time horizon, constraints, and missing variables.

03Recommend

Summarize the assessment, risks, assumptions, and smallest useful next action.

Private AI beta mobile app preview showing a private local AI chat interface

02 Structured output

Built for clear decisions under incomplete context.

  • Direct questions. If the context is thin, it asks for the missing fact before recommending action.
  • Evidence-first reasoning. Screenshots and files are treated as better signal than vague memory.
  • Personal context. Outside private mode, it builds a local memory of themes, preferences, patterns, and decisions.
  • Safety rails. It flags illegal, dangerous, medical, legal, or high-risk requests and keeps the user inside the app's intended scope.

03 Personal Data Net

Memory is deliberate, inspectable, and limited.

Raw private transcripts are not the product. The product is a reviewable local intelligence layer: confirmed preferences, recurring themes, people, projects, decisions, and context the user can inspect or delete.

Local memory objects
Confirmed
User-reviewed facts and preferences.
Tentative
Low-confidence patterns, never treated as facts.
Private mode
Designed for no local memory carryover when private mode is active.

04 Privacy position

Local first. Consent before sharing.

Core chat

Problem solving runs locally where the device supports the model. Basic use does not require uploading private context.

Signup

Email or phone verification creates the beta account and user base. That account is separate from private on-device memory.

Optional data

Research, improvement, sync, and sensor features are opt-in, visible, revocable, and explained before use.

Private mode

When sealed, nothing from that room is remembered after the user exits.

05 Product questions

What makes this different from a general AI assistant?

It is built around a decision workflow. It captures context, asks for missing evidence, names assumptions, warns about risk, and gives a next action instead of producing generic encouragement.

How is this connected to Cast by Stella?

This private beta is a Stella-powered surface for personal decision support. It uses the same principles: local context, inspectable memory, decision structure, and reviewable next actions.

What model does it use?

The first Android builds target small on-device edge models so the experience can stay fast and private. Larger reasoning upgrades can become premium when the hardware supports them.

Can it use screenshots, voice, camera, or watch signals?

The product is designed for optional context. The safe rollout starts with text and user-attached evidence, then adds voice and sensors only with clear consent and settings controls.

06 Private beta

Request access to the private beta.

The Android-first local AI build is in testing. Request access if you want a private decision workspace where context stays organized, inspectable, and reviewable.

Beta access is invite-based. This form collects interest; app access and release notes are handled separately.

Request access

Tell us enough to route the invite and understand the first decision workflow you want supported.