machineai-activation-capacitor

Capacitor (Android) runtime adapter for machineai-activation.

Bridges the native LiteRT-LM Capacitor plugin to the activation runtime contract, so a Capacitor app can activate a local .litertlm model and run on-device inference through the same createMachine / generateText API as every other runtime — instead of hand-writing ~570 LOC of plugin + capability-resolution glue.

Install

npm install machineai-activation-capacitor machineai-activation

This package ships the JavaScript bridge only. The matching native Kotlin plugin (the actual LiteRT-LM execution) has to live in the host Android app — it can't be packaged as an npm dependency. See ACTIVATION_SDK_INTEGRATION_CHECKPOINT.md in the Ingredient Analyzer reference app for the host-app Gradle/Kotlin surface.

Android only. Every entry point throws (or reports unavailable) on web/iOS builds that don't include the native plugin. Guard with isMachineActivationPluginAvailable() before activating.

Usage

Option A — register a global runtime factory

For apps that resolve the runtime lazily from a global (the activation SDK's default lookup):

import { registerCapacitorMachineActivationRuntime } from 'machineai-activation-capacitor';

// In your app entry file, once:
registerCapacitorMachineActivationRuntime();

Option B — pass the runtime explicitly to createMachine

import { createMachine, generateText } from 'machineai-activation';
import {
  createCapacitorMachineActivationRuntime,
  isMachineActivationPluginAvailable,
  pickMachineActivationModel,
} from 'machineai-activation-capacitor';

if (!isMachineActivationPluginAvailable()) {
  throw new Error('Local mode requires an Android build with the native plugin.');
}

const machine = createMachine({
  runtimes: [createCapacitorMachineActivationRuntime()],
});

// Let the user pick a .litertlm package via the native file picker:
const picked = await pickMachineActivationModel();
if (picked) {
  const model = machine.model({ filePath: picked.filePath });
  const { text } = await generateText({ model, prompt: 'Hello on-device!' });
}

What it covers

• Routes .litertlm models to the LiteRT-LM lane (canHandleModel / supportedModelFormats: ['litert-lm']).
• Backend, device, and model-package probing through the native plugin.
• Session creation, complete / completeChat, context-state reporting, vision-readiness probing, diagnostics, abort, and close.
• A native file picker (pickMachineActivationModel) and a plugin-availability guard (isMachineActivationPluginAvailable).

Current limitations

• Android only, .litertlm-first.
• Non-streaming request execution (the native bridge reports supportsStreaming: false today).
• Acceleration telemetry currently reports cpu; richer GPU/NPU reporting is a follow-up.

See CARTRIDGE_SDK_ROADMAP.md and LITERT_LM_ANDROID_NOTES.md in the SDK repo for the deeper LiteRT-LM status.