ModalityDance/PalmClaw

[EMNLP 2026 Demo] "PalmClaw: A Native On-Device Agent Framework for Mobile Phones"

What it solves

PalmClaw provides a way to run a full AI agent framework natively on an Android device. It eliminates the need for a separate PC to manage the agent's logic, memory, and tool execution, allowing the phone to act as the primary host for the agent's operations.

How it works

The framework keeps the agent loop, session management, and context on the device. While the system uses a user-configured LLM API for inference, the actual execution of tasks—such as accessing files, hardware, and personal data—happens locally on the phone. It uses structured device tools and explicit execution boundaries (permissions and confirmations) to ensure security.

Who it’s for

It is designed for mobile AI practitioners, researchers, and users who want a native Android-based AI agent that can interact with their device's data and external messaging channels.

Highlights

  • On-device framework: Manages sessions, memory, and the agent loop directly on Android.
  • Structured device tools: Provides typed operations to access files, hardware, and web automation.
  • External channel integration: Can receive tasks and send responses via Telegram, Discord, Slack, Feishu, WeCom, and Email.
  • Automation capabilities: Supports scheduled work through Cron, heartbeat, and Always-on mode.
  • Extensible: Supports custom LLM providers and MCP servers.

Related

  • Project
  • Project
  • Project
  • Project
  • Project