Droidclaw turns Android devices into autonomous AI agents by implementing a perception-reasoning-action loop. The system reads the device screen via accessibility trees, uses an LLM to decide what actions to take, and executes taps, types, and swipes through ADB. It repeats this cycle until the goal is achieved or a step limit is reached. The agent can also delegate tasks to on-device apps like ChatGPT or Gemini without needing API keys.