Play the video in the post belowwobsoriano/touchpress on GitHub
- Stars
- 28
- Forks
- 2
- Language
- TypeScript
- Last push
- 20 Sep 2026
Post on X
- Views
- 14,748
- Likes
- 196
- Reposts
- 11
Read from GitHub and X on . Counts change daily.
touchpress runs end-to-end tests for mobile apps on the Playwright test runner, driving a booted simulator or emulator through agent-device. Its act method drives the app from an instruction in English, and it runs on either a language model or an evaluation model such as Jev. With Jev, each step is a choice question with a calibrated distribution over the moves the current screen offers. Jev never writes text, so it types only the values the instruction puts in double quotes, and a pass it gives below an 80% chance is looked at again rather than trusted.
Soriano reports that the same test on the same loop took 29 seconds and $0.003 with Jev against 37 seconds and $0.07 with Haiku 4.5. The AI docs break down a sign-in flow: Jev took six moves with calls of 120 to 450 milliseconds, while Claude Haiku 4.5 took the same six moves with calls of one to three seconds. The README warns that the codebase was largely written by an LLM and is highly experimental.
