Introduction

Mobile Agent is an open-source AI agent built specifically for Android devices. The agent runtime, memory, skills, tools, and local workspace all live inside the app itself — there is no separate companion server, no cloud backend, and nothing that must run on your computer.

What it does

Mobile Agent turns your phone into an autonomous assistant that can:

• Chat with large language models, either on-device (fully offline) or from cloud providers
• Connect to MCP servers to reach external tools like Notion, GitHub, Slack, Stripe, and dozens more
• Read and write files in its own workspace and, with Android's Storage Access Framework, in external folders
• Remember useful facts between conversations with persistent memory
• Package repeatable instructions into reusable skills
• Run entirely on your schedule — including in the background while the app is closed
• Execute commands live in Termux and stream the output back into the conversation
• Extend itself with JavaScript plugins

How it works

Everything is local-first. Your conversations, schedules, memory, skills, agents, and the model catalog are stored on-device in a SQLite database managed by the app. When you choose an on-device model, inference happens right on your phone's hardware with no network connection at all.

When you connect a cloud provider such as OpenAI or Anthropic, only the messages for that conversation are sent to the provider you chose. Nothing is ever sent through a Mobile Agent proxy — you own your keys and your data.

Key features

• On-device AI models — LiteRT-LM models (Gemma, Qwen, Phi, DeepSeek, and more) that run entirely offline in a background download manager.
• Cloud model providers — OpenAI, Anthropic, Google, xAI, OpenRouter, Ollama, and any OpenAI-compatible or Anthropic-compatible endpoint.
• MCP support — the Model Context Protocol with a preset catalog of 50+ servers, plus custom servers over HTTP or SSE, including full OAuth.
• Permission-first tools — every file, terminal, and device action is gated by Android permissions and per-tool approval settings.
• Reusable skills — markdown-defined instructions that activate automatically when the task matches.
• Persistent memory — a declarative fact store that survives across conversations.
• Custom agents — plan/build agents, subagents, and per-agent tool permissions.
• Scheduler — cron-based jobs that run in the background with exact alarms and notifications.
• Termux terminal — live command streaming from the AI to the terminal and back.
• File workspace — project-style file access with list, read, write, grep, glob, and edit tools.
• Multi-modal — send images to capable models and ask the model to generate images when the provider supports it.
• Open source — MIT licensed, with an actively maintained repository and release pipeline.

Design principles

1. Local first. The default stack runs on your device. Cloud services are opt-in per conversation.
2. Permission-first. The agent can only do what you explicitly allow, one approval at a time.
3. Understandable. Every model, server, skill, and tool has a clear name, description, and status.
4. Packaged, not pasted. Skills, agents, and plugins can be exported as plain markdown/JavaScript and imported elsewhere.

Installation

To get started, download the latest APK from the GitHub Releases page and install it on your Android device. See Installation for details, or jump straight to the Quick Start guide.