Skip to content

Cross-Domain Product Stories ​

These end-to-end scenarios show how subsystem boundaries work together in real product flows.

Scenario 1: Creating a Character and Running a LocalAgent ​

  1. Character Selection: The user creates or selects a Character in a Realm-backed surface.
  2. Source Generation: Realm issues the structured Character Source definition.
  3. Local Instantiation: Runtime instantiates an account-scoped LocalAgent from that definition on the user's machine.
  4. Authorized Projection: Nimi Home or an app receives only the authorized LocalAgent projection through the SDK.
  5. Private State Isolation: Runtime keeps conversation sessions, Cognition-backed memory routing, inference loadouts, readiness state, and provider credentials inside its own boundary.

The host surface never mints character identities, and Realm never executes the LocalAgent's local inference turns.

Scenario 2: Continuing a Conversation Across Surfaces ​

  1. Session Creation: Runtime opens a conversation and returns an explicit, unique Conversation Anchor.
  2. Multi-Surface Rendering: The Desktop host renders committed turns and manages lightweight local UI state.
  3. Embodiment Attachment: Avatar binds a 2D or 3D character model to the same authorized Conversation Anchor.
  4. Single Continuity Authority: Runtime remains the sole source of truth for conversation continuity; neither surface copies local chat logs to fabricate recovery state.

Avatar manages gaze tracking and physical motion locally, while Runtime projects dialogue text, speech synthesis, and presentation timing.

Scenario 3: Launching a Scaffolded Application ​

  1. Protected Launch: Nimi Home starts the application through a local protected launch sequence.
  2. Intent Submission: The application submits typed operations and the session-scoped agentHandle through accepted SDK APIs.
  3. Session Verification: Runtime derives the user account, application identity, and allowed operations directly from the protected host session.
  4. Least-Privilege Return: The application receives a validated, typed result—never Realm JWTs, provider API keys, internal Runtime tokens, or account-wide agent rosters.

Direct SDK integrations and scaffolded standalone apps follow the same architectural contract. These boundary rules are platform invariants, not user-toggled settings.

NOTE

Long-term memory is isolated per agent in Cognition and accessed through Runtime. Knowledge V1 currently has no active public client interface.

Source Basis ​

Documentation for Nimi — the installable, open-source, local-first personal AI product.