Clash Jev
A research Clash Royale bot that uses computer vision and Jev to choose gameplay actions on a real Android device.
Added to Jevfast
How it uses Jev
A computer-vision pipeline reads Clash Royale state from an Android device. Jev first chooses a strategy from a fixed Choice list. If that strategy plays a card, later Choices select one of four hand slots and a legal board square. The runner applies the selected action after physical checks and logs the questions, probabilities, and action source.
What Jev decides
Game strategyExample answers · not a recorded Jev response · Source ↗
Question 1 · strategy
YOUR APP
INSTRUCTION
Choose a strategy, including waiting for elixir, defending, counter-pushing, or pushing a lane.
STATE
Current game briefing, clock, elixir, hand-card costs and profiles, estimated opponent resources, unit positions, and tower health from computer vision.
JEV · CHOICE
- save_elixir
- hold_elixir_for_threat
- defend_left
- defend_right
- counter_push_left
- counter_push_right
- push_left
- push_right
- defend_centre
- build_push
- split_push
- cycle
App workflow
- Observe the match
Stream the Android game screen and form a measured state snapshot with computer vision.
- Ask for a strategy
Send the changed state to Jev and record the strategy Choice.
- Choose and apply a play
If play is selected, ask Jev for a hand card and legal placement square, then validate and tap the device.
- Log and repeat
Save decisions and synchronized video; wait for a changed state before another loop step.
Why it is interesting
The bot exposes the model’s decisions in an embodied game loop: broad strategy choices, measured game state, sequential card and placement questions, and synchronized logs/video make author-reported behavior inspectable. API failure is labeled as a fallback policy.