Examples

Runnable examples for every package, from a plain Dart script to a live camera app.

Every package in this SDK ships an example that creates a real task, runs inference, and closes it. The examples are built in continuous integration and run on a device, so one that stops compiling breaks the build.

Continuous integration#

Two commands run the same code paths in CI:

devenv shell test:examples
devenv shell test:android-device

test:examples runs the Dart examples on the host against a real native runtime. test:android-device drives the mp_text native plugin fixture on an attached Android device, and test:device-demo launches the Flutter demo on one.

Dart examples#

These run with dart run from the repository root. Each downloads its model once, verifies the SHA-256 digest, and caches it under the system temporary directory or MP_EXAMPLE_CACHE.

Example	Demonstrates
`examples/cli/bin/core_model_assets.dart`	Model assets, image and audio containers, timestamp rules, and typed failures
`examples/cli/bin/vision_face_detection.dart`	Face detection over a real photograph with bounding boxes and keypoints
`examples/cli/bin/vision_hand_landmarks.dart`	Hand landmarks, handedness, and world coordinates on the same photograph
`examples/cli/bin/text_language_detection.dart`	Language identification across four languages
`examples/cli/bin/text_tasks.dart`	Language detection plus text embedding and cosine similarity
`examples/cli/bin/audio_classification.dart`	Clip classification and the native streaming contract
`examples/cli/bin/genai_llm.dart`	LLM session creation, streamed chunks, and cancellation

Run one:

dart run examples/cli/bin/vision_face_detection.dart

The native examples need a MediaPipe runtime. A Flutter build resolves one through the native asset hook; a standalone dart run reads a local .mp-sdk directory:

devenv shell native:build
dart run examples/cli/bin/vision_face_detection.dart

Flutter device demo#

examples/device_demo is a live camera playground: it streams frames through pose and face landmarkers, counts exercise reps, and projects 3D accessories onto the tracked face. It is the end-to-end test for the camera path: frame conversion, monotonic timestamps, latest-frame backpressure, and deterministic cleanup on shutdown.

cd examples/device_demo
repo-flutter run

The demo requests the camera permission itself and downloads its digest-pinned models on first launch, so no manual setup is required. A browser build runs in a synthetic-preview mode without a camera.

The repository also carries a second, device-only entry point, lib/device_telemetry.dart, which streams a face detector while reporting device motion, processed frames, and dropped frames:

cd examples/device_demo
repo-flutter run -t lib/device_telemetry.dart

What the examples prove#

A task is created, used, and closed without leaking native or JavaScript resources.
Timestamps advance monotonically across a live stream.
A slow model drops stale frames instead of building an unbounded queue.
An unsupported platform documents itself through MpStatus.unimplemented rather than failing vaguely.

Read the live streams guide for the scheduling rules these examples rely on, or the models guide for how to supply a model safely in production.