VIOLOOP

Most AI assistants stop at the answer: you still copy the text, switch apps, fill in the fields and click send. VIOLOOP is pitched at that last stretch of manual work. The maker says the device reads what is on your screen, understands the context, suggests the next step and can then open apps, navigate, fill in forms, handle files and run scripts across several stages of a task.

Seeing the screen instead of being told about it

Rather than relying on app-specific APIs or plug-ins, VIOLOOP takes in the PC's video output over HDMI 2.0 at up to 4K/30Hz, with USB-C video input also supported and configurations of up to two displays envisioned. It works with Windows, macOS and Linux. Because it presents itself to the computer as a standard keyboard and mouse, the maker says ordinary desktop and web apps, including older software without APIs, can be operated from the screen. Where possible it uses more direct routes such as scripts or system accessibility features, falling back to keyboard and mouse input when those aren't available. File operations and script execution rely on companion apps for PC and smartphone.

The target is repetitive, rule-based work: moving data between apps, entering forms, updating records, working down lists, sorting messages and drafting replies, and organizing files. A screen-aware suggestion feature surfaces proposed actions near the cursor, accepted with Tab. Users can describe tasks in natural language or demonstrate a workflow once and save it for reuse; the device is said to retain what it learns about procedures and app-to-app handoffs on the device itself.

Away Mode and physical approval

In Away Mode, according to the campaign, a task can continue after you leave your desk, even with the PC locked if needed, with smartphone notifications when the job finishes or a decision is required, and status checks and approvals available from the mobile app. Consequential actions such as sending, confirming, overwriting, paying and deleting pause until approved via a physical button or an authenticated smartphone. That gate runs on a dedicated STM32H563 security MCU, separate from the chip handling the AI.

Local processing by default

• RK3576 octa-core processor plus dedicated AI accelerator, 26 TOPS combined
• 13GB memory (8GB LPDDR4X and 5GB 3D-stacked DRAM)
• 128GB eMMC 5.1 storage
• Runs Qwen 8B Q4-class models on-device
• AMOLED touchscreen, dual-band Wi-Fi, Bluetooth 5.2, voice control
• Approx. 90×92×80mm, approx. 450g

The maker states that raw screen video is not continuously streamed to the cloud. In fully local mode, task content stays on the device; for jobs needing larger models, only the necessary context is passed to an external model depending on how it is used, and users can bring their own local models or API keys.