Eyes and hands on a machine, even before the OS starts
Bootscry is building a shared interface for an AI agent to observe and act on a physical machine. Video KVM, serial consoles, cameras, and input devices can each become a driver. Today the GL.iNet Comet is the working reference for screen, keyboard, and mouse control, including before an operating system starts. Other channels are on the roadmap.
Who it is for
Systems engineers, datacenter operators, and AI agent developers who need a path to machine control when remote software access fails. We are developing multi-agent coordination with echo.cc.
The proof: real boot-menu catch
On a physical mini PC, the Comet delivered 119 Delete key taps during a graceful reboot and reached the AMI setup utility. We recorded the KVM video, exited without saving firmware changes, and verified the machine returned to its login prompt.
One interface, several kinds of eyes and hands
A driver declares the observations and actions it can provide. Video KVM supplies a screen and USB input; serial could supply exact console text; a camera could observe the physical environment and eventually provide audio; HID-only hardware supplies input but must be paired with live feedback. Shared safety rules are planned across every channel.
How to try it
Contact the maintainer about a pilot on supported hardware. The code is private while the Comet reference implementation is being prepared for public release.
Hardware demonstration & capture status
Bootscry operates on physical hardware. Our reference benchmark executed on the psycho mini PC connected to a GL.iNet Comet KVM.
Documented physical capture sequence
- Reboot: A graceful SSH command restarted the target. This Comet has no ATX board.
- Hotkey loop: The KVM began sending Delete immediately after the reboot command returned. It sent 119 taps over 35 seconds, without using OCR to stop the loop.
- Result: The recorded screen showed the AMI Aptio Setup Utility. The operator inspected the frame after the fixed key loop ended.
- Clean exit: Escape opened the "Exit Without Saving" dialog; Enter confirmed it. A later frame showed the Linux login prompt with no user logged in.
Measured evidence at a glance
These figures come from one physical Comet test bench and a separate operator account of a NAS recovery. Each result has its own scope and limits.
| Measurement | Measured result | Evidence document |
|---|---|---|
| Kernel mouse positioning | Within 1/32767 in target kernel; within 1 px in X across 17 raw and 26 graphical checks | docs/evidence/mouse-accuracy.md |
| Keystroke reliability | 25/25 exact at full speed; 25/25 in slow mode (50/50 total) | docs/evidence/typing.md |
| BIOS OCR phrase match | RapidOCR score 0.98; Tesseract 0.94 on 3 frames and 30 phrases | docs/evidence/bios-ocr.md |
| Strict verbatim phrases | Local VLM 31/32; Tesseract 22/32; RapidOCR 16/32 on 4 frames | docs/evidence/vlm-vs-ocr.md |
| NAS recovery, operator account | The operator's post-mortem records a 180-second watchdog cycle and 21 of 23 checks passed, with two warnings. It does not attribute each action to an agent. | docs/evidence/nas-recovery.md |