title: Capability boundaries description: Separate AIRI source capabilities from configured sessions, native prerequisites, and evidence that a hosted feature actually works.
Capability boundaries
The public showcase is a source guide. It does not establish an active AI character, a configured account, or access to a visitor’s microphone.
A source file proves that code exists. Configuration proves that selections exist. Runtime acceptance requires an observed result on the intended application surface.
flowchart LR Source[Source implementation] --> Prerequisites[Required assets, services, and platform] Prerequisites --> Configuration[Valid configuration] Configuration --> Operation[Actual operation] Operation --> Evidence[Observed result]
Skipping a step changes the claim. A settings screenshot cannot replace a model request, a rendered asset, or an audible output check.
What each capability needs
| Capability | What the source contains | What a real session still needs |
|---|---|---|
| Text generation | Consciousness selection and a conversation runtime | Reachable provider, available model, and required credentials or account access |
| Speech recognition | Hearing module and several provider paths | Audio input, permission, platform capability, and the selected transcription subsystem |
| Speech output | Speech module and speech-provider integrations | A real provider selection, required model and voice, and working audio output |
| Visible character | Display-model imports and renderer packages | Compatible assets, graphics support, and a successfully rendered stage |
| Local Ollama chat | An adapter with a localhost endpoint | A separately installed server and model weights |
| Local Kokoro TTS | WASM and WebGPU model variants | Required model files and a supported backend for the selected variant |
| Apple Speech | An Electron-native provider | macOS desktop runtime and a successful native availability check |
| Messaging and games | Integration source and stage-side modules | Separate processes, accounts, permissions, and reachable external systems |
| Hosted account services | API and Auth source plus a local backend stack | Deployed services, secrets, databases, and service readiness |
These requirements are not a minimum hardware specification. This wiki publishes no unmeasured latency, memory, GPU, or device-support guarantee.
Claims this guide does not make
- Every feature runs on the public showcase host.
- A provider catalog entry contains a configured provider account.
- A model name identifies a locally installed model.
- Browser speech recognition always runs offline.
- Every local inference path requires WebGPU.
- A desktop-only provider works in a browser or on Linux.
- Imported files always render correctly.
- A messaging or game settings page starts its external process.
- A complete long-term memory product exists because a settings route exists.
- A source build proves runtime, accessibility, or production readiness.
Known incomplete and conditional areas
The reviewed long-term memory settings page renders a WIP component. It is not a usable long-term memory feature merely because navigation reaches it.
The core-agent guide states that its Realtime transport is not implemented. Its existing generation and streaming contracts do not establish a continuous realtime session adapter.
The Minecraft guide labels its current Mineflayer integration as a migration target and warns against untrusted public servers.
Provider availability checks remain narrower than product guarantees. Browser Web Speech checks API presence. Apple Speech checks the desktop runtime, macOS, and native availability.
Kokoro explicitly supports both WASM and WebGPU variants. The selected WebGPU model fails its availability check without WebGPU support.
How to state a result honestly
A useful result names the application surface, source revision, prerequisite, action, and observation. It also names anything that remained unavailable.
For example, an imported model visible after reload proves more than an import notification. It still does not prove voice, inference, or another device.
A source-derived workflow is a procedure, not a completed experiment. Use workflows to choose the next operation and operations to inspect failures.
Source context
The immutable revision and unexercised areas are recorded in provenance.
packages/stage-ui/src/stores/modules/consciousness.ts, lines 117–119: configured selection predicate.packages/stage-pages/src/pages/settings/modules/memory-long-term.vue: incomplete memory surface.packages/core-agent/README.md, lines 41–47: protocol support and absent Realtime transport.packages/stage-ui/src/libs/providers/providers/kokoro-local/index.ts, lines 42–50: selected-model requirement.packages/stage-ui/src/workers/kokoro/constants.ts, lines 9–71: available execution platforms.integrations/minecraft/README.md, lines 5–17: migration and trusted-server limit.
Navigation: index · providers-and-models · services-and-integrations