Mobile-Agent

Mobile-Agent is a family of multimodal GUI agents from Alibaba’s X-PLUG team that can see a phone screen and operate apps the way a person would. The open-source GitHub project has collected more than 9,300 stars, and the family keeps growing, with updates continuing into 2026.

Overview

Given a natural-language instruction such as booking, searching or posting, a Mobile-Agent perceives the current screen, reasons about what to do next, and executes taps, swipes and text input on an Android device. Later versions add multi-agent collaboration, with planning, decision and reflection roles that catch mistakes and recover from unexpected screens, plus memory of past interactions. The project includes the Mobile-AgentBench evaluation suite for measuring agent performance across real apps, and it has become one of the reference codebases for research on phone-use agents. The codebase also ships trajectory-recording tooling, so teams can capture full agent runs, study failure modes in detail, and compare model backbones on identical tasks.

Key Features

• End-to-end GUI automation on Android from natural-language instructions
• Multimodal perception that grounds UI elements from screenshots without app-specific APIs
• Multi-agent designs with planning, decision-making and self-reflection to reduce errors
• Experience memory so the agent improves across repeated tasks
• Mobile-AgentBench benchmark for evaluating mobile agents on real applications
• Support for strong vision-language models as the reasoning backbone

💳 Pricing

Mobile-Agent is free and open source under the MIT License. You provide the device or emulator and any vision-language model API access the agent uses.

Who Is It For

Mobile-Agent is for researchers and developers building phone automation, accessibility tools or app-testing agents. It gives you working reference implementations of modern GUI-agent techniques, from single-agent control loops to reflective multi-agent systems, ready to study and extend.