MeowClaw

MeowClaw: an on-device Android agent that operates your phone with a virtual mouse and keyboard. Native Kotlin, Needle + Cactus on-device models, or any OpenAI-compatible API.

MeowClaw is an open-source Android agent that operates your phone for you. It's a native Kotlin + Jetpack Compose app. Quick actions and chat run on-device, and it controls other apps through a virtual mouse and keyboard.

How it works request ──► Needle 3 (on-device, ~30 ms) ──► quick action (open app, call, SMS, alarm, timer, volume…) │ multi-step / unsure ▼ Fast path ──► taps named elements, types searches — no model call │ anything else ▼ Brain: on-device LLM (Cactus) or any OpenAI-compatible API │ ▼ Screen loop: read screen (+ screenshot for vision models) → pick one step → act → repeat Needle 3 (Cactus Compute). A 35 MB tool-calling model. It turns simple requests into device actions offline, in tens of milliseconds. Its results are checked against your words: timer durations and alarm times are re-read from the request, and "open X" must really say "open X". Fast path. Goals like "open settings then open display" or "open chrome and search for weather in london" run deterministically. The agent taps the named element or types into the search box without a model call. Brain. For chat and open-ended tasks, pick one: On-device via the Cactus engine: Gemma 4 E2B: vision, tools, audio Qwen3.5 0.8B / 2B: vision, tools LFM2.5 350M / 1.2B: tools LFM2.5-VL 1.6B, LFM2-VL 450M: vision FunctionGemma 270M API: any OpenAI-compatible endpoint (DeepSeek, OpenRouter, Groq, NVIDIA NIM, Ollama, llama.cpp / LM Studio on your LAN). Vision. Vision models receive a screenshot each step. The agent's own overlay is hidden while it's captured. Native tool calling. On-device tool models answer with constrained function calls, so small models stay on the rails. Virtual mouse & keyboard The agent's "hands" try the best available backend for every action:

Accessibility (default) Input method ADB via Shizuku Tap, double tap, long press, drag, wheel-scroll dispatched gestures — input tap/swipe/draganddrop Type text ACTION_SET_TEXT Android 13+ accessibility IME, or the optional MeowClaw Keyboard input text Keys & shortcuts (Enter, Tab, arrows, Ctrl+A…) IME editor action real key events input keyevent / keycombination While the agent is in control, a cursor glides to each target with a click ripple and an action label, and the screen edge glows. Both are drawn in a touch-transparent accessibility overlay, so no "display over other apps" permission is needed.

Features Chat and Agent modes, voice input, spoken replies, chat history Skill memory: successful step sequences replay without the model Task history with full traces and token counts Telegram remote control. The first chat to message the bot becomes its only owner. Neubrutalist UI in light and dark themes Private by default: Cactus cloud telemetry and handoff are compiled out; keys stay on the device

Setup Install the APK and open MeowClaw. Hands: enable MeowClaw Screen Control in Accessibility settings. If Android shows "Restricted setting", open Settings → Apps → MeowClaw → ⋮ → Allow restricted settings first. Brain: download Needle (35 MB). Then either: download an on-device model (Gemma 4 E2B or Qwen3.5 2B for multi-step tasks), or add an API key. For free cloud use: OpenRouter + openai/gpt-oss-120b:free. Optional: Shizuku for ADB-level fallbacks Modify system settings for brightness control