This skill is built for use inside a development organization (e.g. a game studio's internal tooling). This page covers what it does and does not do with respect to licensing, network access, and data.

The core runtime dependencies are permissive and compatible with the MIT-licensed skill. (Optional OCR pulls its own dependencies into a separate venv from requirements-ocr.txt -- paddleocr/paddlepaddle and their transitive deps, both Apache-2.0; audit them separately with pip-licenses in .venv-ocr if required.)

No strong-copyleft (GPL) dependency; one weak-copyleft component (FFmpeg, LGPL-2.1) is bundled in the opencv wheel -- see the note above, it only matters if you redistribute the binary. Verify in your environment with pip-licenses (not a runtime dependency) or pip show <pkg>.

• The core makes no network access. It opens no outbound connections, sends no telemetry, and phones nothing home -- fully usable on an air-gapped machine. (OCR is optional and its first run downloads models unless you vendor them offline; see OCR models below.)
• The only files it writes are the captures you ask for (PNG/MP4 at the path you give) and a small recovery ledger (see below).

• Captured frames are written as local files where --out/--record point. Nothing is uploaded.
• The crash-recovery ledger at %LOCALAPPDATA%\capture\parked.json stores window handles, placement and style flags, the target window's class and process identity, and the owning capture process's identity -- no window content (no pixels).
• Captures can contain whatever is on screen. Treat the output files with the same sensitivity as a screenshot.

A --title that matches more than one window is refused, not silently resolved, so an ambiguous match can't grab a sensitive window you didn't mean. Disambiguate with --hwnd (authoritative) or --exact (exact-title match). This is enforced on every resolution path (single capture, --wait, and --loop streaming).

For locked, reproducible installs, generate a hash-pinned requirements file on the target platform and install with --require-hashes:

pip download -r requirements.txt -d wheels/        # fetch the exact wheels
pip hash wheels/*                                  # print --hash lines
# or, with pip-tools:
pip-compile --generate-hashes requirements.txt -o requirements-lock.txt
pip install --require-hashes -r requirements-lock.txt

Hashes are platform- and Python-version-specific (Windows / CPython 3.14 here), so generate them in the environment you deploy to.

OCR (optional, a separate Python 3.13 subprocess) uses PaddleOCR. The OCR venv brings its own dependencies (paddleocr, paddlepaddle, and their transitive deps) -- audit them separately with pip-licenses in .venv-ocr if your policy requires it; they are isolated from the core. The first OCR run downloads the PP-OCR models to %USERPROFILE%\.paddlex. For air-gapped or corporate deployment, vendor the models offline -- download them once on a connected machine and copy the .paddlex model cache to the target, so the worker never fetches anything at runtime.

Development-process windows -- editors, tools, dev builds. Not anti-cheat-protected live games (PrintWindow returns black under exclusive fullscreen, and window manipulation can trip anti-cheat).