Why Hardware Integration Matters
OtherU requires a trusted execution and interaction path. The Framework Desktop provides high-performance local inference while JetKVM creates a deterministic, out-of-band control channel for any connected computer — no software installation required on the target machine.
Framework Desktop
OtherU Core is designed around the Framework Desktop with an AMD Ryzen AI Max+ 395 (Strix Halo) processor.
| Component | Detail |
|---|---|
| CPU | AMD Ryzen AI Max+ 395 (Strix Halo, 16-core Zen 5) |
| GPU | AMD Radeon 8060S iGPU (gfx1151, RDNA3.5) |
| Memory | 128GB unified (CPU + GPU share the same pool) |
| Storage | NVMe SSD |
| OS | Linux with ROCm stack |
Why unified memory matters: The iGPU accesses all 128GB of system RAM directly. Models that would require a high-end discrete GPU — including 46GB+ LLMs — run on this machine without any discrete GPU card.
JetKVM Role
JetKVM is a compact hardware device that provides:
- HDMI capture — real-time visual context at full resolution via WebRTC
- USB HID injection — deterministic mouse and keyboard control of any computer
- Out-of-band management — operates regardless of target OS state (BIOS, lock screen, crashed OS)
The target machine sees only a USB keyboard and mouse — no drivers or agents needed.
Integration Pattern
Coordinator
│
▼ typed interaction intent (click x,y / type "text" / press Enter)
Hardware Bridge (port 8005)
│
▼ coordinate validation + inverse transform (vision space → real screen)
JetKVM
│
├── /screenshot → WebRTC frame capture
├── /mouse/move → USB HID mouse movement
├── /mouse/click → USB HID click
└── /keyboard → USB HID keypress / type
Screenshot Pipeline
- Bridge captures raw frame from JetKVM WebRTC stream
- Letterboxes the capture to the vision model's input resolution
- Returns
image_base64+letterbox_scale+letterbox_padding+mouse_position - Fara-7B sees the letterboxed image and decides an action with coordinates in that space
- Coordinator inverse-transforms coordinates back to real screen space before HID injection
Mouse Position Tracking
- Optical flow: Lucas-Kanade tracker initialized at screen center
- Confidence scoring: 0.0–1.0; falls back to dead-reckoning if tracking drift exceeds threshold
- Cursor hint: current cursor position injected into Fara's observation context so it doesn't need to locate the cursor visually
Display Configuration
Always use the display's native resolution EDID. Custom low-resolution EDIDs may lack fallback timing modes, causing the target GPU output to go dark. Software letterboxing in the bridge pipeline achieves the same result without touching the EDID.
# Reset to native EDID via bridge
curl -X POST http://127.0.0.1:8005/edid \
-H 'Content-Type: application/json' \
-d '{"edid_name": "T749-fHD720"}'
Operator Guidance
- Pin JetKVM firmware version before any production rollout
- Test capture latency and HID round-trip after bridge updates
- Hardware bridge source is baked into its Docker image — apply patches via
docker cpthen restart - Set
MAX_RECORD_DURATIONappropriately for your use case (default is conservative)