Native fetch returns a page.
Levirge returns the thing you wanted.
Same URL given to two tools — the host's native web fetch, then Levirge fetch, unchanged. Output quoted verbatim, dated, reproducible. This table grows every time we hit one during normal work.
| URL | Native web fetch | Levirge fetch | Route | Date |
|---|---|---|---|---|
| youtube.com/watch?v=SvOlwP9-3yc | Succeeded — page shell only: "the actual transcript is not included … only the YouTube page header and footer information are visible" | Title, channel, views, length, videoId, full description with timestamped chapters — 16.2 KB of markdown | direct | 10 Aug 2026 |
| reddit.com/r/selfhosted/top/ | Refused: "Claude Code is unable to fetch from www.reddit.com" | The live feed as markdown — 35.9 KB | stealth | 10 Aug 2026 |
| x.com/deno_land | "HTTP 402 Payment Required. The response body was not retrieved." | Profile and recent posts, pinned post first | direct | 10 Aug 2026 |
"Native" is Claude Code's built-in WebFetch, captured on the date shown — other clients differ in detail, not outcome. Rows are only added when the difference reproduces on capture day: a LinkedIn company page read fine in both tools the same day, so it isn't listed.
Read the routes honestly
direct means no stealth browser was involved — the difference is per-content-type extraction, not access: a watch page becomes its transcript, an API body becomes the fields you asked for. stealth means the source refuses ordinary tools and a full browser opened the page.
Read vs query
Native web fetch only accepts URLs that already appeared in the conversation or a search result. Levirge fetch also accepts URLs the agent constructs itself — an API path with parameters it reasoned out — so the agent can query, not just read.