arXiv's API doesn't send CORS headers, so a genuinely browser-only tool can't talk to it directly - this routes through a small proxy (a cPanel PHP passthrough, not a third-party service) that hits arXiv's API and parses the result. The proxy doesn't log bibliographic content, caches by ID hash (30 days for versioned IDs since those are immutable, 7 days otherwise), and rate-limits per IP.
The actual bug this fixes: most converters emit eprint = {arXiv:1706.03762}, but biblatex's \eprint macro prepends "arXiv:" itself at render time - so the rendered citation reads "arXiv:arXiv:1706.03762." The acmart LaTeX class hit exactly this (it's issue #382 on their tracker). The fix is emitting the bare ID in eprint and putting the prefix in archivePrefix instead.
Other specifics: citekeys come out as vaswani2017attention rather than the raw numeric ID some converters use (arxiv2bibtex.org keys on the literal 1706.03762, which breaks alpha.bst labels and makes \cite{} unreadable). Handles old-style IDs (math/0506203, hep-th/9901001) and new-style with version suffixes (1706.03762v2) preserved if included. Inline math in titles like $O(n \log n)$ is detected and passed through unescaped. Defaults to @ misc for true preprints but exposes a per-row override - community convention is genuinely split here, some prefer @online with eprinttype={arxiv}, some use @article - switch per entry as needed. If the paper's also published with a DOI, the DOI tool is the cleaner path for that entry.