← all posts site ↗

building my own pocket agent computer

saw a tiny device from a company called pamir ai, the lapis one. it is an 85 millimeter cube of linux, 227 grams, with a battery, a monochrome memory lcd like the one in the playdate, and a fingerprint sensor. the pitch is that it is a pocket agent computer, an always-on box you carry around, plug into anything, and let a cloud model drive it. early bird price is around 360 dollars and the first 140 units ship in october. the part that got me is the agent kvm port, it captures the target machine's screen and pretends to be a keyboard and mouse, so the agent can operate literally any computer with zero software installed on it, down to the bios.

the kvm trick is two separate tricks combined. first the box uses a usb controller that can flip roles, instead of talking to devices it becomes a device, so the target pc just sees a keyboard and mouse appear out of nowhere. this only works on raspberry pi class boards, the cheap x86 mini pcs have host-only usb so they can never pull it off. second, a cheap hdmi capture dongle feeds the screen back into the box as video frames. the agent looks at the screen, types, looks again, corrects itself. that is the whole magic, see and touch with nothing installed on the target.

i want one badly but the diy route gets most of the way there for maybe half the price. a pi 5 with a battery hat, a small case and display, and the capture dongle. i do not even need the npu because the brains live in the cloud anyway. the image that gets me is a portable box with an led strip and a capacitive panel, something i can plug into a hotel tv or a random desktop and have my agent just be there, watching and typing. very appealing little machine.