MediaPipe

Google's on-device machine learning toolkit gives app developers prebuilt vision, text, and audio tasks that run locally on Android, iOS, the web, desktop Python, and embedded boards. Vision tasks include face detection and landmarks, hand and pose landmarks, gesture recognition, holistic tracking, object detection, classification, and segmentation; text and audio tasks cover classification, language detection, and embeddings. Each task pairs a pretrained model with a high-level API, Model Maker retrains several of them on your own data, and MediaPipe Studio provides browser demos for quick evaluation. Underneath, a C++ framework wires calculators together into graphs that pass packets of data, so you can assemble custom real-time pipelines with GPU acceleration where the platform supports it. Packages exist for pip, npm, Android, and iOS. Input data stays on the device, though the Tasks APIs send usage metrics to Google that apps must disclose. The code is Apache 2.0 and free for commercial use; version 1.0.0 arrived in July 2026. It suits mobile and web developers who need live tracking without a server.