Claude Royale

Claude Royale

  ▶️ Watch: Live 2v2 match with Claude as my teammate

A harness for AI agents to play Clash Royale autonomously.

What is this?

Claude Royale is an open framework that lets AI systems play Clash Royale through screen capture and input automation. It handles the game interface, coordinate mapping, and action execution - you provide the AI decision-making.

The included implementation uses Claude Code with a multi-agent architecture: a Commander agent manages menus and game flow while spawning 3 parallel Player agents that work together to play each match.

Current progress: 1000+ trophies (Spell Valley - Arena 5)

Live tested: 12+ hours streamed on Twitch with Claude playing live

Autonomy: Claude played 14 games in a row autonomously, with minor human intervention only to close pop-ups and open chests

How it works

Why 3 agents? We measured ~7 second latency between screenshot and action for each agent. Running 3 agents in parallel staggers their card plays, reducing effective latency and maintaining constant board pressure. As AI inference speeds improve, fewer agents may be needed.

Learnings

Slow Opening Cards
Early matches were lost because Claude took too long to play the first card - opponents would rush and take a tower before Claude reacted. We added an auto-opener that plays the first card at 5 seconds and a second card at 8 seconds after hitting the battle button. This establishes early tempo and forces the opponent to react rather than freely push.

7-Second Latency
We added debug logging to the sub-agents and discovered that the round-trip from screenshot → vision analysis → decision → tool call takes roughly 7 seconds. This is why we use 3 parallel agents playing the same match - their natural stagger means a card gets played…