Index / PalmClaw · updated Oct 5, 2026
PalmClaw
ModalityDance/PalmClaw · healthy · rank 32 of 50 by stars
An academic (EMNLP 2026 demo) OpenClaw-inspired agent framework that runs natively on Android phones with no PC required. It pairs on-device session/memory/tool management with user-configured LLM APIs and a permissions-first security model.
Facts
Repository
Runtime
Posture
Community
Security breakdown
Composite 78 / 100 · how these are scored
higher is safer
higher is safer
higher is safer
higher is safer
higher is riskier
Model access
Read from the repository, not written by a model · 34 files examined
1 provider · custom endpoint
Pinned models
gpt-5.4-mini released 2026-03-17 — 7 months old (from the public model catalogue)
Pin last edited 40 days ago
Several defaults are pinned; the project picks one per provider rather than shipping a single default.
Evidence
Decision
Why choose PalmClaw over OpenClaw?
Why choose this
- Runs natively on Android with no PC or server needed
- Structured device tools expose mobile-specific capabilities
- Permissions, confirmations, and workspace boundaries built in
Tradeoffs
- Inference depends on external LLM APIs, not local models
- Much smaller ecosystem and community than OpenClaw
- Early-stage (v0.3.x) with academic-demo maturity
Best fit
- Users who want an agent running entirely on an Android phone
- Researchers evaluating mobile-native agent architectures
- Developers building on structured mobile device tools
Avoid if
- You need iOS or desktop support
- You want a fully offline model (inference requires an external LLM API)
- You need a mature plugin ecosystem and production stability
README and release data give solid evidence of architecture and activity, but Reddit matches are unrelated (mouse-grip discussions) and web sentiment is about the broader OpenClaw ecosystem, so community and sentiment scores are low-confidence.
AI layer reviewed Oct 5, 2026 · how this is written
Star activity
1,166 stars today
Overview
PalmClaw is an open-source, EMNLP 2026 demo agent framework explicitly inspired by OpenClaw but re-architected to run natively on Android phones. Instead of requiring a PC or server, it keeps sessions, context management, memory, skills, tools, and the agent loop entirely on the device, while delegating model inference to a user-configured LLM API. It ships as a downloadable APK with 769 tests and active development through v0.3.1.
Its distinguishing features are mobile-native structured device tools and a control-oriented security model: actions are gated by schemas, permissions, confirmations, and workspace boundaries. Recent commits show work on MCP client lifecycle hardening, always-on reliability, and provider endpoint compatibility, suggesting a push toward robustness.
Compared to OpenClaw, PalmClaw trades ecosystem breadth and multi-channel reach for true on-phone operation and a low-friction install. It is best viewed as an early-stage, research-grade mobile alternative rather than a production-ready replacement.