References¶
Reasoning Core References¶
- OpenClaw: https://github.com/openclaw/openclaw
- PicoClaw: https://github.com/sipeed/picoclaw
Use these projects to study agent-loop structure, session handling, prompt/context layering, skill loading, memory boundaries, tool policy, and event logging. TouchPilot should adapt these patterns to a smaller 100% local Android runtime.
Android Control References¶
- MobileClaw: https://github.com/MobileClaw/MobileClaw
- ClawMobile: https://github.com/ClawMobile/ClawMobile
Use these projects to study observe-act loops, screenshot and UI-tree handling, Android action tools, trace capture, workflow replay, and demonstration-to-skill patterns.
UI/Event UX Reference¶
- wende/mobileclaw: https://github.com/wende/mobileclaw
Use this project to study mobile chat UX for visible agent work: tool-call cards, running/success/error states, reasoning state, diffs, traces, and event streams.
Local Runtime References¶
- LiteRT: https://ai.google.dev/edge/litert
- ExecuTorch: https://github.com/pytorch/executorch
- llama.cpp: https://github.com/ggml-org/llama.cpp
Use LiteRT first for compact local routing, classification, ranking, OCR-adjacent models, and other small Android-friendly inference. Keep ExecuTorch and llama.cpp as future local LLM/VLM runtime candidates.
Notes¶
TouchPilot should borrow patterns, not copy identity. The product direction is a 100% local hybrid Android AI agent: local model reasoning where ambiguity exists, deterministic routing where correctness matters, local skills and workflows for repeatable knowledge, and policy/approvals for safety.