ai_tracker 0.1.0

On-device observability (cost, latency, tokens, errors) for any AI/LLM call from a Flutter app. Local-first & zero-config by default. Optional opt-in sync to your own backend. Pure-Dart core, works on [...]

ai_tracker

On-device observability for every AI/LLM call in your Flutter app — cost, latency, tokens, errors. Local-first & zero-config by default. Optional sync only to your own backend. Pure-Dart core: Android, iOS, web, desktop, no platform channels for tracking.

---

Build & run (this repo)

This package uses drift code generation. Generate once before compiling:

flutter pub get
dart run build_runner build --delete-conflicting-outputs

Run the demo (no API keys needed — it simulates calls):

cd example
flutter create --platforms=android,ios,web .   # generate native host projects
flutter pub get
flutter run                                     # mobile/desktop

To run the demo on web, fetch drift's WASM assets first:

bash ../platform_setup/web/fetch_web_assets.sh web
flutter run -d chrome

Run tests:

flutter test

Platform setup for sync + web: local-only tracking needs no platform config anywhere. Enabling sync (Android/iOS background) or running on web requires the steps in PLATFORM_SETUP.md.

Version constraints in pubspec.yaml are conservative. If pub get complains, run flutter pub upgrade --major-versions and re-run build_runner.

---

(a) Fully local, zero config — under 60 seconds

import 'package:ai_tracker/ai_tracker.dart';

void main() async {
  WidgetsFlutterBinding.ensureInitialized();
  await AiTracker.init();               // 1. one line
  runApp(const MyApp());
}

// 2. wrap any AI call — result is returned unchanged
final reply = await AiTracker.track(
  provider: 'openai',
  model: 'gpt-4o',
  operation: () => myOpenAiCall(prompt),
  extractUsage: (r) => AiUsage(
    inputTokens: r.promptTokens, outputTokens: r.completionTokens),
);

// 3. see it
Navigator.push(context, MaterialPageRoute(
  builder: (_) => const Scaffold(body: AiTrackerDashboard())));

No server. No INTERNET permission required. Data never leaves the device.

Streaming calls

final stream = AiTracker.trackStream(
  provider: 'openai',
  model: 'gpt-4o',
  operation: () => myOpenAiStream(prompt),   // Stream<Chunk>
  extractUsage: (chunks) => AiUsage(
    inputTokens: chunks.last.promptTokens ?? 0,
    outputTokens: chunks.last.completionTokens ?? 0),
);
await for (final chunk in stream) { /* your UI */ }
// captures time-to-first-token AND total latency automatically

On-device inference (TFLite / ML Kit / Gemini Nano)

Same call — you just supply the token count, since there's no HTTP response to parse. These calls make no network request, so proxy-based tools can't see them; ai_tracker can.

await AiTracker.track(
  provider: 'gemini-nano',
  model: 'nano-2',
  operationType: AiOperationType.onDeviceInference,
  operation: () => geminiNano.generate(prompt),
  extractUsage: (r) => AiUsage(outputTokens: r.tokenCount),
);

---

(b) Optional: aggregate across all your users — 3 lines

Point the SDK at your own backend. Off unless you turn it on.

await AiTracker.init(
  syncEnabled: true,
  syncEndpoint: 'https://myapp.com/api/ai-events',  // your server, nobody else's
);

Events batch (N events or T minutes, whichever first), upload in the background (survive app kill), retry with exponential backoff when offline, and carry no prompt or response text — metrics only. A ready-to-host FastAPI receiver is in reference_backend/.

Platform note: background uploads use workmanager on Android/iOS (Android minimum period is 15 min; iOS runs opportunistically via BGTaskScheduler). Web and desktop have no OS scheduler, so there sync runs via the in-process timer while the app is alive. Full config in PLATFORM_SETUP.md.

---

Export a shareable HTML report (no backend, ever)

dart run ai_tracker:report            # writes ai_report.html
dart run ai_tracker:report <db-path>  # if the db isn't at the default location

---

Configuration

AiTracker.init(...)	default	meaning
syncEnabled	false	master switch for upload — off by default
syncEndpoint	null	your server URL (required when syncEnabled)
syncIntervalMinutes	15	flush cadence
syncBatchSize	50	flush early once this many pile up
pricing	built-ins	Map<String, ModelPrice> overrides

---

License

MIT (add a LICENSE file before publishing).