Public API surface (@runanywhere/web)

The root package intentionally exports a small Swift-shaped surface. Prefer a flat RunAnywhere.* method whenever Swift exposes a flat method for the same operation; only reach for a namespace (RunAnywhere.solutions, RunAnywhere.pluginLoader) when Swift itself has a namespace property there, or when backend/package internals need a lower-level handle. Example/app code should stick to root methods and avoid @runanywhere/web/internal.

See ../../AGENTS.md for the initialization sequence this facade sits on top of.

Method inventory

Model lifecycle/registry RunAnywhere.loadModel, unloadModel, currentModel, componentLifecycleSnapshot, listModels, queryModels, getModel, downloadedModels, downloadModel, importModel.

LLM / structured output / tool calling RunAnywhere.generate, generateStream, cancelGeneration, generateStructured, generateStructuredStream, extractStructuredOutput, generateWithTools.

Speech / VLM / Voice Agent / RAG RunAnywhere.transcribe, transcribeStream, synthesize, synthesizeStream, speak, stopSynthesis, stopSpeaking, detectVoiceActivity, streamVAD, resetVAD, processImage, processImageStream, cancelVLMGeneration, initializeVoiceAgent, processVoiceTurn, streamVoiceAgent, ragCreatePipeline, ragIngest, ragQuery, etc.

Quick-start shape

import { RunAnywhere, SDKEnvironment } from '@runanywhere/web';
import { LlamaCPP } from '@runanywhere/web-llamacpp';

await RunAnywhere.initialize({
  environment: SDKEnvironment.SDK_ENVIRONMENT_DEVELOPMENT,
});
await LlamaCPP.register({ acceleration: 'auto' });

const stream = await RunAnywhere.generateStream({
  prompt: 'Write a haiku about local AI.',
  maxTokens: 128,
});

for await (const token of stream.stream) {
  console.log(token);
}