google-ai-edge/mediapipe

What it is

MediaPipe provides on-device machine learning components for real-time audio, vision, text and generative AI workloads. It has two layers: MediaPipe Tasks, cross-platform APIs and libraries plus pretrained models for common tasks such as object detection, audio classification, text classification, LLM inference and semantic retrieval; and MediaPipe Framework, a lower-level graph engine for building custom perception pipelines out of packets, graphs and calculators framework overview.

It targets developers shipping ML features to Android, iOS, web, desktop, edge hardware and IoT, rather than researchers training models. The repository is C++ under Apache-2.0, and the Solutions suite is built on the same open-source code so solutions can be adapted.

Getting started

Start with the solution guide matching your task, such as vision object detection, audio classification or LLM inference. Then follow the setup guide for your target platform:

• Android setup
• iOS setup
• Python setup
• Web setup

Before integrating, you can try the browser-based demos or inspect and benchmark models in MediaPipe Studio. The README's image-based hero and demo pages are the intended entry point; concrete API calls live in the platform setup guides rather than in the repository README.

Custom pipeline work

If the premade Tasks are insufficient, install MediaPipe Framework and build example applications in C++, Android or iOS. Framework work requires understanding three concepts first: packets, graphs and calculators. This is the more involved path and is aimed at teams that need custom pipeline topology or low-level control.

Constraints and data handling

MediaPipe Tasks process input data such as images, video and text on device and do not send that input to Google servers, which permits work on data that should stay local. The Tasks APIs do send performance and utilization metrics to Google, and the README states that app developers are responsible for obtaining informed consent from users for that processing privacy notice. Platform setup differs by target; there is no single install command covering all platforms.

HotRepo's take

MediaPipe is a large, long-running Google project, with 37,681 stars and 6,224 forks as of the 2026-11-14 daily C++ listing, where it gained 13 stars and reached position 13. That single snapshot shows continuing attention in the C++ ecosystem, but one day of data cannot establish a trend. For readers, the relevant signal is maturity and breadth: the same framework serves mobile, web, desktop and edge targets, so it is a reasonable default when you want pretrained on-device tasks without assembling a pipeline from scratch. Teams that need a custom graph still face the steeper Framework learning curve.