System requirements

Requirements for the complete product depend on the workflow: markup and classic CV are substantially lighter than model-backed CV, long Capture sessions and Screph Code. The sections below separate confirmed technical boundaries from recommendations that do not yet have a clean-machine compatibility matrix.

1. Platform and compatibility status

The current target platform is Windows 10/11 x64. Fresh packaged screph.exe, pro_agent.exe and QtWebEngineProcess.exe binaries use the AMD64 PE machine type; 32-bit Windows is not a supported target. Linux and macOS remain roadmap items, and running the source there is not equivalent to product support.

Alpha evidence boundary. The 1.0.9 artifact manifest does not record architecture, a minimum Windows build or compatibility-matrix results. The QtIFW installer also performs no separate OS-version preflight; the available packaged smoke applies to the current build host, and no clean-machine Windows 10/11 matrix was found in release evidence. “Windows 10/11 x64” is therefore the support target, not a certification for every Windows 10 update.

2. CPU and memory

The current installer/runtime has no validated minimum CPU-core or RAM threshold, so the site does not invent values such as “8 GB minimum.” Actual load depends on frame size, FPS, timeline length, the number of canvas presentations, selected CV models, the LLM runtime and parallel Screph Code processes.

  • For markup, projects and classic OpenCV, start without the ML component.
  • For long Capture sessions, use disk_images; ram_images stores frames in memory and may require lower FPS or shorter sessions as usage grows.
  • For model-backed CV, start with small models and frames; caches may retain models in RAM/VRAM until cleared or the process exits.

3. GPU and CUDA

A discrete GPU is not required for the main product. YOLO defaults to use_gpu=false and explicitly selects CPU when CUDA is disabled or unavailable. An NVIDIA GPU with a compatible CUDA/PyTorch runtime accelerates heavy models, but does not replace the ml.runtime and model-weight installation.

OmniParser device-control boundary. Its visible GPU checkbox also defaults off and is persisted, but the current CV runner does not read it: run_omniparser() keeps device=None, leaving device selection to the underlying runtime. Until that consumer is connected, the checkbox is not a guarantee that CPU or GPU mode is being forced.

Large YOLO, OmniParser or SAM tasks may be slow on CPU. On CUDA out-of-memory, reduce the image or model size; YOLO can explicitly disable GPU use. Screph publishes no universal minimum VRAM because it depends on the model and input.

4. Disk space

  • Full 1.0.9 configuration: about 3.9 GiB of uncompressed default-selected QtIFW components.
  • Optional packaged ML component: about 1.36 GiB more.
  • Separate managed ML pack: up to 6.5 GB installed; its network archive is estimated at about 334 MB.

The installer's dependency page calculates installer.requiredDiskSpace(), adds the managed-ML estimate and blocks continuation unless the target drive has the estimate plus a 512 MB reserve. This is the most accurate check for the selected components. It does not reserve space for future projects, Capture frames, model weights, download caches, logs or Assistant/Screph Code workspaces; keep separate headroom for those.

5. Runtime and embedded web engine

A normal packaged installation does not require a preinstalled system Python: the main application and managed dependency payload provide their own runtime files. The Dependencies tab can reuse compatible global tools only after a probe; the absence of global Python is not an installation error.

Screph Code uses bundled QtWebEngine/Monaco: the full payload includes QtWebEngineProcess.exe, resources and PySide6 bindings. Microsoft WebView2 is not a separate Screph prerequisite. Browser-based device login and external links still require an available system browser.

6. Devices, network and permissions

  • Screen, markup and classic CV run with ordinary current-user permissions; the default target is %LOCALAPPDATA%\Screph.
  • Camera and voice require the relevant devices and Windows permissions; a stream URL requires a supported OpenCV backend and codec.
  • GUI action execution requires a configured Arduino HID, legacy serial or FakerInput Virtual HID; driver installation uses UAC, while Arduino flashing modifies the external board.
  • Core local workflows can run offline after runtimes and models are provisioned. Updates, archives, cloud LLM or speech, account access and support use the network separately.

Windows, UAC and devices →
Runtime components and repair →