AI mobile automation framework
Android and iOS
Open-source framework for AI agents that control Android and iOS devices. LLM-agnostic. Multi-agent architecture. Built for developers.
1import asyncio2from droidrun import DroidAgent, DroidConfig3 4async def main():5 config = DroidConfig()6 agent = DroidAgent(7 goal="Open Twitter and post 'Hello from droidrun!'",8 config=config,9 )10 result = await agent.run()11 12asyncio.run(main())Works with any LLM
Gemini 
OpenAI 
Anthropic 
Ollama 
DeepSeek 
OpenRouter 
Qwen Gemini OpenAI Anthropic Ollama DeepSeek OpenRouter Qwen Gemini 2.5 Pro/Flash · OpenAI GPT-5.4 / o3 · Claude Opus 4.6 / Sonnet 4.6 · DeepSeek R2 · Qwen 3.5 · any Ollama model
watch it in action
See droidrun control real Android devices with natural language.
automate anything on mobile
Reliably and effortlessly. Three agent archetypes for every workflow.
Task Agents
Automate complex multi-app workflows through natural language
Testing Agents
Write mobile tests in plain English — no Appium, no XPath
Research Agents
Extract structured data from any mobile app
built in the open
Join thousands of developers building the future of mobile automation.
“Replaced our entire Appium suite in a weekend. 400 test cases, all natural language now.”
“The A11Y tree approach is genius. 2KB vs 1MB screenshots changed everything for our CI pipeline.”
“Cross-app automation is the killer feature. Nothing else can chain WhatsApp to Sheets to Gmail.”
“Action replay saved us thousands in LLM costs. Record once, replay at 100ms/step.”
“pip install and running in 30 seconds. Compared to 2 days setting up Appium.”
“Typed extraction with Pydantic models is clean. App store scraping that returns real objects.”
“Replaced our entire Appium suite in a weekend. 400 test cases, all natural language now.”
“The A11Y tree approach is genius. 2KB vs 1MB screenshots changed everything for our CI pipeline.”
“Cross-app automation is the killer feature. Nothing else can chain WhatsApp to Sheets to Gmail.”
“Action replay saved us thousands in LLM costs. Record once, replay at 100ms/step.”
“pip install and running in 30 seconds. Compared to 2 days setting up Appium.”
“Typed extraction with Pydantic models is clean. App store scraping that returns real objects.”
AndroidWorld benchmark
droidrun achieves state-of-the-art performance on AndroidWorld with 91.4% success rate across 116 diverse tasks.
Accessibility tree vs. screenshots
Traditional mobile automation uses screenshots and pixel matching. We use the accessibility tree — 500x smaller payloads.
everything you need to ship
Six primitives that cover the full mobile automation lifecycle.
LLM Agnostic
Swap providers with one line. Gemini, OpenAI, Anthropic, Ollama, DeepSeek, OpenRouter, Qwen and more.
Vision Mode
Screenshot-based fallback when the accessibility tree is insufficient. Combine both for maximum accuracy.
Reasoning Mode
Multi-step planning with chain-of-thought. The agent reasons about UI state before acting.
Structured Output
Extract typed data with Pydantic models. Get real objects, not raw strings.
Custom Tools
Register your own tools the agent can call. Extend capabilities beyond built-in actions.
App Cards
Pre-built interaction patterns for popular apps. Skip the boilerplate, start automating.
device setup
Use your own Android or iOS devices with the open-source Portal app. Secure, real-time, fully wireless. No root or jailbreak required.
droidrun setup
droidrun ping
[ok] device paired -- readystart local. scale to cloud.
Same agent framework. Run locally on your device or remote on mobilerun.
droidrun
- • pip install
- • Local devices
- • Full control
mobilerun
- • Hosted devices
- • Parallel agents
- • Stealth mode
Run automation locally using your own Android devices or emulators. Install with pip, inspect every line of code, and stay fully in control — no cloud setup, no vendor lock-in.
When you're ready to go beyond local, mobilerun offers hosted devices, unlimited parallel agents, global proxies, and stealth-mode execution. Production-grade infrastructure, zero maintenance.
| Legacy | droidrun | mobilerun | |
|---|---|---|---|
| Core Features | |||
| Can read screen content | ✓ | ✓ | ✓ |
| Can interact with device | ✓ | ✓ | ✓ |
| Continues to work when UI changes | ✗ | ✓ | ✓ |
| Extract data using natural language | ✗ | ✓ | ✓ |
| Framework Standards | |||
| Easy to setup | ✗ | ◐ | ✓ |
| Self-healing | ✗ | ✓ | ✓ |
| Open source | ✓ | ✓ | ✗ |
| Infrastructure | |||
| Stealth & anti-detection | ✗ | ✗ | ✓ |
| Hosted device fleet | ✗ | ✗ | ✓ |
| No setup required | ✗ | ✗ | ✓ |
from zero to agent in 60 seconds
Three commands. One device. Full automation.
$ pip install droidrun
Successfully installed droidrun-0.5.8
$ droidrun setup
Portal is installed and accessible. You're good to go!
$ droidrun run "Open Settings and check battery level"
[agent] planning steps...
[step 1/2] tap "Settings" on home screen
[step 2/2] navigate to "Battery"
[done] task completed in 3.8s -- 2 actions, 0 retriesfrequently asked questions
How is droidrun different from using Claude/GPT with computer use?+
Generic LLMs send full screenshots to the model (~1MB per frame). droidrun uses the accessibility tree (~2KB) — that’s 500x smaller, 10x faster, and dramatically cheaper. Plus droidrun has Action Replay (record once, replay without LLM costs). A generic agent explores blindly. droidrun knows what it’s doing.
How is droidrun different from Appium?+
Appium requires writing test scripts with exact element selectors that break when the UI changes. droidrun is AI-native — describe what you want in natural language and the agent figures it out. It self-heals when UI changes, works on real devices, and supports both Android and iOS. Users report replacing entire Appium suites in a weekend.
Which LLMs does droidrun support?+
droidrun is fully LLM-agnostic. Works with OpenAI (GPT-4, GPT-4o), Anthropic (Claude Sonnet, Opus), Google (Gemini 2, Gemini 3), DeepSeek, Ollama (local models), OpenRouter, and Qwen. Switch models with a single config change.
What is the difference between droidrun and mobilerun?+
droidrun is the open-source framework (the agent brain) — install with pip, run on your own devices, full control. mobilerun is the cloud platform (the device infrastructure) — hosted real Android and iOS phones with stealth mode, residential proxies, and fleet management. Start local with droidrun, scale to the cloud with mobilerun. Same agent protocol.
Is droidrun free?+
Yes — droidrun is open-source under the MIT license. pip install droidrun and you’re running in 30 seconds. For hosted cloud devices with stealth mode and residential proxies, mobilerun plans start at $5/month with a free tier for OpenClaw users.
start building today.
Clone the repo, run your first automation, and scale to production on mobilerun.ai when you're ready.