PhysiClaw is an AI agent with a physical body: a gantry arm moves a stylus across a phone's screen while a camera watches it, so the agent operates the phone the way a person does. It needs no APIs, no OAuth, no ADB cable, and nothing installed on the phone.
The approach targets apps that expose no public API and that flag simulated input such as desktop automation or ADB. To the phone, the stylus is a finger on the screen, so the agent reaches nearly any app. The trade-off is speed: each action takes a few seconds.
PhysiClaw has its own iPhone running its own WhatsApp account, and you message it like a contact with tasks such as ordering takeout, buying groceries, booking a ride, paying a bill, or replying to a message. Its runtime wakes the agent when a scheduled task is due or a new message lights up the screen. The agent unlocks the phone, does the task, pauses for confirmation on steps such as payments, replies with the result, saves its memory, and exits until next woken.
The agent never runs commands on your computer or touches your files. You choose the LLM provider and model it runs on. Both the CAD-as-code hardware and the agent software are open source, so you can buy the parts and assemble your own.
Features
- Physical operation: a stylus performs taps, swipes, and scrolls while a camera reads the screen
- No integration access: works without APIs, OAuth, jailbreaks, or software installed on the phone
- Chat interface: you message the agent on its own phone, like texting a contact
- Action verification: the camera checks after every action whether it landed and the job got done
- Scheduled tasks: the runtime wakes the agent for cron jobs and new messages
- Confirmation on payments: pauses for your approval before sensitive steps
- Contained scope: keeps only its memory and scheduled jobs, and never touches your computer or files
- Open hardware: CAD-as-code arm, stylus, and camera mount you can assemble yourself
