Mano-P
An open-source GUI vision-language-action agent for edge devices that automates on-screen tasks purely from visual input, running locally on an Apple Silicon Mac.
Mano-P is an open-source GUI-aware agent model designed for edge devices, combining a vision-language-action model with skills for autonomous, multi-step GUI automation. It ranks first on a specialized GUI-automation benchmark and is being released in phases: first the GUI-automation skills for agent tools such as OpenClaw or Claude Code, then the local inference models and SDK for building custom skills, and eventually the training, pruning, and quantization methods behind the models themselves.
Its GUI-VLA models run inference directly on an Apple M4 Mac mini or MacBook with 32GB or more of RAM, or on a compute stick connected over USB, so all screen data and automation stay on the device and are not sent to an external server.
It is released under the Apache License 2.0.