Everyone is obsessed with one question- If AI is really that good, can it pick my phone and do stuff automatically.
In my opinion, why not…AI is absolutely capable of it and it’s not a future tech fantasy anymore. It’s happening right now, right in front of us.
AI can already order food for you just by asking.
It can book a ride when you’re running late and still getting ready (that’s me most of the times)
It can handle sales, customer service, billing, customer journeys, and almost anything you want to automate.
This is exactly what Droidrun is doing.
Bottleneck to Breakthrough: Droidrun’s journey
We at Droidrun started working on this project around 7–8 months ago. The technology came a long way- from making the agent do random clicks to making it do everything precisely (even if the app UI changes).
Earlier, the agent struggled with text tasks, unstable UI interpretation, and limited app understanding. We addressed this by creating a dedicated text-manipulation agent that programmatically constructs and replaces text using structured accessibility data. We then strengthened contextual awareness by stabilizing screen-state detection, filtering visual noise, tracking screen changes, and extracting app-specific knowledge for more accurate system understanding.
To make the framework predictable, we added transparent action reporting and expanded the action space with long-press, targeted typing, system buttons, swipe gestures, and app launching. We reinforced memory across the prompt flow and refined the system through extensive testing, turning it into a reliable and capable mobile automation framework.
Earlier, the agent occasionally got things right; now it can decide when to run, when to click, where to click, and what the user is asking for. To enable natural-language execution, we built a complete framework that abstracts core actions like clicking, scrolling, typing, screenshots, and UI understanding. Then, we taught the agent to interpret commands and the mobile environment like a real user.
Repo-https://github.com/droidrun/droidrun
Moving to the cloud — Mobilerun
After testing and refining the framework, now we are taking the next step towards building the cloud. Mobilerun is like having a virtual phone that runs fully in the cloud. It keeps all your sensitive data encrypted and safe, and it allows you to replay, pause, and audit every action. Everything happens with transparency and control.
Why did we keep it open-source?
The tech behind Mobile AI agents is continuously evolving. The primary intent of open-sourcing Droidrun was to turn this AI capability into a public foundation that the entire industry can build on, improve together, and innovate freely.
As a developer, you can treat it as a building block. You can extend it, modify the code, or transform it into something entirely new.
Check out our Github repo here- https://github.com/droidrun/droidrun
How are we ensuring privacy and security
Users can use proxy. It makes it possible to connect to the app from different geographical areas.
Credentials of the users are not sent to LLM. They are saved into the file as key-value pairs and the agent (droidrun) fetches them by using a script. LLM does not see these credentials. So, no leaks will happen.
Users don’t use their real phone. So apps cannot understand the real identity of the user if users won’t use their private emails and nicknames.
What’s Next?
We are continuously working on the cloud, making it more reliable, more stable and more stealth. And this is only the beginning. We would love to see what users can build with our framework and later on the cloud. Start building, experiment, and explore. If you get stuck, reach out to us on Discord.
Moving forward, we will keep building together toward the kind of agent we all imagine. The era of Mobile AI Agents is here, and we want everyone to see it, experience it, and use it for themselves.
Keep Building
Top comments (0)