betto_onnxrt

ONNX Runtime for Dart and Flutter — binary delivery via the native-assets build hook, a generalised inference API, and model-download infrastructure.

Wraps ONNX Runtime v1.22.0. Targets macOS, Linux, Windows, Android, and iOS. Web is excluded (no FFI in the browser).

betto_onnxrt has no Flutter dependency and works in CLI, server, and Flutter contexts. Add betto_onnxrt_ios as an additional dependency only when targeting iOS.

Add to pubspec.yaml:

dependencies:
  betto_onnxrt: ^0.1.0-dev.2
  betto_onnxrt_ios: ^0.1.0-dev.2 # Flutter + iOS targets only

Run dart pub get (or flutter pub get). The native-assets build hook downloads and verifies the ORT binary for your target platform automatically — no manual binary management needed.

import 'package:betto_onnxrt/betto_onnxrt.dart';

final runtime = OnnxRuntime.load();
final session = runtime.createSession('/path/to/model.onnx');
final outputs = session.run({'input': inputTensor});
session.dispose();

See the specification for the full API.

packages/betto_onnxrt/example/magika/ is a standalone Dart CLI tool that detects file types using Google's Magika v3.3 ONNX model. It demonstrates end-to-end use of OnnxRuntime, OnnxSession, and ModelDownloader with a real-world model, and its output mirrors the Python Magika --json format.

cd packages/betto_onnxrt/example/magika

# Run directly (JIT — no build step needed):
dart run bin/magika.dart <file>

# Or compile to a native binary:
dart build cli
./build/cli/macos_arm64/bundle/bin/magika <file>

The model is downloaded and SHA-256 verified on first run; no manual setup is required.

All build, test, and quality targets are available from the repo root via make. The Makefile composes per-package .mk fragments.

make              # full quality gate (format, analyze, test, coverage, doc)
make pre_commit   # subset run before committing
make test         # unit tests
make macos_test   # on-device integration tests (macOS)
make ios_test     # iOS simulator integration tests
make android_test # Android emulator integration tests

See CLAUDE.md for the full command reference and architectural notes aimed at AI-assisted development.

Pre-release (0.1.0-dev.2).

See the v0 roadmap for remaining work before the first stable release.

Apache 2.0. See LICENSE.