PalmClaw
An open-source agent framework that runs entirely on an Android phone, keeping sessions, memory, skills and tools on the device while model inference comes from a user-configured LLM API.
PalmClaw is an open-source agent framework inspired by OpenClaw, built to run natively on an Android phone without a separate PC. Sessions, context management, memory, skills, tools, and the agent loop all run on the device, while the model itself is called through a user-configured LLM API.
It can access mobile capabilities such as files, personal data, hardware, web, and communication through typed device tools, gated by explicit schema checks, Android permissions, user confirmations, and workspace-path boundaries before any action executes. Conversation history, tool traces, attachments, session workspaces, and reusable memory persist across turns, and the app can receive tasks from local chat or connected channels and run scheduled work through cron jobs, a heartbeat, or an always-on mode, using built-in or custom LLM providers, MCP servers, and installed skills.
It is written in Kotlin and is described in an accompanying paper accepted to the EMNLP 2026 System Demonstrations track.