Mobile Jev
Choose bounded actions on a real Android device, execute through Mobilerun, and verify fresh device state.
What Jev decides
Whether to open an app, tap, type an exact supplied span, scroll, navigate, wait, finish, or report that the task is blocked.
Mobile Jev observes indexed controls and installed apps, asks Jev for an operation and compatible target, validates the selected branch, and executes it through the Mobilerun API.
- Jev role
- Select the next bounded Android operation and a compatible target from fresh device observations.
- Input
- The user goal, installed-app inventory, indexed controls and bounds, current device observation, and bounded action history.
- Output
- Choice distributions for the operation and compatible speculative targets in one request.
- Decision frequency
- Once per observe-decide-act cycle on the configured Android device.
- Fallback
- Reject stale or incompatible targets, stop after uncertain mutation failures, and verify readable field values and demo postconditions before continuing.
How it works
Overview
Mobile Jev is a standalone Android agent with a local React studio, CLI, execution traces, and request-level timing. Each loop observes the device, builds compatible actions and targets, asks Jev to select among them, validates the branch and target freshness, and then executes through Mobilerun.
Text is copied from exact spans in the user goal rather than generated as arbitrary prose. The included dark-theme demo verifies the actual Android setting after the run. A separate recorded Uber flow reaches payment selection in about 21 seconds over nine actions, but does not demonstrate a completed booking. Running the project requires both TypeSafe and Mobilerun access.
- Language
- JavaScript
- License
- MIT
- Author
- DroidRun
Sources
Related projects
typesafe-computer-use
A native macOS computer-use loop built from local OCR, accessibility data, and bounded Jev actions.
What Jev decides
Which computer action, visible item, site, or off-screen accessibility control to use next.
Jev Autopilot
A Jev autopilot flies a simulated drone across a generated Three.js city and lands on a target pad.
What Jev decides
Throttle, yaw, pitch, roll, whether to commit to landing, and whether to cut the motors.
Distill
Route coding-agent work, reasoning effort, and context retention through bounded Jev decisions.
What Jev decides
Which configured model and effort should handle a call, what evidence deserves attention, whether another attempt is useful, and what context should be retained.