Droidrun

Today I shipped mobile-jev, and I want to explain why it's more than a cool demo. For two years, "computer use" agents have all worked the same way: screenshot in, huge LLM reasons in text, action parsed out. It works, but it's slow, expensive, and can still return something malformed. Then TypeSafe dropped Jev, the first "System One" model. Instead of generating text, it returns a typed, calibrated decision in under half a second, for a fraction of a cent. No screenshots, no chain of thought, just structured decisions. I wanted to test that idea on the hardest surface there is: a real mobile app, on a real phone. That's mobile-jev. Jev makes every decision. Mobilerun (our execution API) makes it happen on an actual Android device. No ADB. No separate model bolted on for typing or reasoning. Jev reads the screen's UI state, picks one typed action (tap, type, scroll, open app, wait, done), and Mobilerun executes it on real hardware. The demo: Jev opens Uber and books a route from SFO to the Golden Gate Bridge, reaching payment selection in about 21 seconds across 9 actions. No prompt engineering, no retries. Is this the answer to computer use agents? Too early to say. But decisions instead of generated text is a real shift, and I think it's worth watching closely. Open source, check it out: https://lnkd.in/eJRcjuXQ

Today I shipped mobile-jev, and I want to explain why it's more than a cool demo. For two years, "computer use" agents have all worked the same way: screenshot in, huge LLM reasons in text, action parsed out. It works, but it's slow, expensive, and can still return something malformed. Then TypeSafe dropped Jev, the first "System One" model. Instead of generating text, it returns a typed, calibrated decision in under half a second, for a fraction of a cent. No screenshots, no chain of thought, just structured decisions. I wanted to test that idea on the hardest surface there is: a real mobile app, on a real phone. That's mobile-jev. Jev makes every decision. Mobilerun (our execution API) makes it happen on an actual Android device. No ADB. No separate model bolted on for typing or reasoning. Jev reads the screen's UI state, picks one typed action (tap, type, scroll, open app, wait, done), and Mobilerun executes it on real hardware. The demo: Jev opens Uber and books a route from SFO to the Golden Gate Bridge, reaching payment selection in about 21 seconds across 9 actions. No prompt engineering, no retries. Is this the answer to computer use agents? Too early to say. But decisions instead of generated text is a real shift, and I think it's worth watching closely. Open source, check it out: https://lnkd.in/eJRcjuXQ

Codex can now control real iOS and Android phones. 📱 Introducing https://mobile-harness.com, an open-source agent harness powered by mobilerun-core, giving agents one unified API and a consistent control path across local and cloud mobile devices.

Just playing around on our platform Droidrun

One problem I often run into when developing mobile apps is this: everything works in the simulator, but not on a real device. There are many reasons for that, including different hardware, different OS versions, different screen sizes, and different device-specific behaviors. Then users open the app on their own phones, and suddenly one or two things break. That got me thinking more seriously about how to automate mobile testing with AI agents. I came across mobilerun (https://lnkd.in/d4j7gi_7) by Droidrun, an open-source framework for controlling Android and iOS devices with LLM agents. It gives agents mobile-native tools to inspect UI state, understand screenshots, tap, swipe, type, plan multi-step workflows, and return results through a CLI or Python API. The first setup was smooth: I got it working on my newer phone running iOS 18. But when I tried using my old phone (an iPhone 8), mobilerun didn’t work out of the box. That part mattered to me, because I specifically want to test apps on older OS versions and smaller screens. So I tweaked mobilerun a bit, downloaded a different Xcode bundle, and eventually got it working. Now my agent can control this old phone too. For the demo, I asked Claude to create a poem in the Notes app 😄 (the video was also edited by Claude using Remotion skills). This feels like a real breakthrough. I can imagine a future of agentic QA where AI agents actively hunt for bugs, log issues automatically, and another agent picks them up and fixes them. All without us touching anything. #iOSDevelopment #OpenSource #ClaudeCode

I won the German Startup Award 2026 as Newcomer of the Year 🤘🏻 Still feels unreal. I’m incredibly honored to receive this recognition from the German startup ecosystem. Thank you to Startup-Verband, Verena Pausder, the jury, and everyone who made this award possible. This award is called “Newcomer of the Year,” but nothing about building a company happens alone. It reflects the work of the team, the trust of our customers, the support of our investors and advisors and all the people who believed before there was much to point to. Huge congratulations to the other winners: Julie Lepique, Prof. Dr. Pia Wülfing, Judith Dada, Simon Schmincke, Ali Abderrahmane, Jarek Kutylowski, Ann-Kristin Achleitner, Ruth Bosse. 🏆 Grateful for the recognition. Now back to building. 🤖📲 Niels Schmidt, Peter Lächner, Nikolai Dück