Read a page

curl -X POST https://api.marrow.navii.online/v1/scrape \
  -H "Authorization: Bearer mrw_xxx" -H "Content-Type: application/json" \
  -d '{"url":"https://en.wikipedia.org/wiki/Web_scraping","formats":["markdown"]}'

The page comes back in data.data.body.

Parameters

field	default	what it does
url	—	required. Missing it is 400 url-required.
formats	["markdown"]	markdown is the only format served. Anything else is refused by name — see When a read stops. ⚠️ The singular format is not accepted; it used to mean something else (Obsidian frontmatter) and a silently-ignored spelling would have changed what you got.
depth	"shallow"	"deep" runs a real browser: JavaScript executes, hidden sections are revealed. Slower, and the only way to read a page that draws itself client-side.
maxAge	86400000 (24h)	Milliseconds. How old a stored capture may be before we go and fetch again. 0 forces a live capture. The freshness default, in full.
max_pages	1	Follow the page’s own “next” control up to this many pages, 1–25. Declaring more than 1 is the request to walk — you do not also need depth:"deep". Outside the range is 400 max-pages-invalid.
max_tokens	none	A ceiling on what comes back. See below.
timeout	30000	Milliseconds allowed for each page load. The whole read has its own 90-second deadline regardless.

Billing. One read is one credit; a walked read is one credit per page actually fetched, so a max_pages: 10 call costs at most 10 and bills what it got. A read served from cache is 0.1. A read that stops is not billed at all.

Capping the size

🚨 max_pages bounds what we fetch. It does not bound what you are handed. Ten pages of a link index is ten page loads on our side and one large document on yours, and the two numbers are not related: an index page is mostly hrefs, an article is mostly prose.

max_tokens is the second ceiling. It is estimated, not tokenised by any particular model — ASCII at 4 characters per token, everything else at 2 — so treat it as a bound, not a count.

-d '{"url":"https://news.ycombinator.com/","max_pages":2,"max_tokens":3000}'

What it cost you: tokenUsage

Every read reports its size, budget or no budget. With no max_tokens, you get the measurement and nothing else:

"tokenUsage": { "rawTokens": 11003, "used": 11003, "truncated": false }

Add a budget and four more fields appear, describing the cut:

"tokenUsage": {
  "rawTokens": 28819, "used": 3000, "budget": 3000, "truncated": true,
  "mode": "blocks", "blocksIncluded": 37, "blocksTotal": 402
}

field	always?	means
rawTokens	✅	the size of the document before any ceiling. Read it once without a budget to find out whether a budget is worth setting at all.
used	✅	what you were actually handed. Equal to rawTokens when nothing was cut.
truncated	✅	whether anything was dropped.
budget	only with max_tokens	the ceiling you asked for. ⚠️ Absent means you did not ask for one — it is never reported as 0, because 0 would mean you asked for nothing.
mode	only with max_tokens	how it was cut. blocks = whole blocks were selected by score, so you got complete pieces. text = there were no blocks to choose from, so the body itself was cut at the ceiling.
blocksIncluded / blocksTotal	only with max_tokens	in blocks mode, how much of the document made it. In text mode both are 0 — ⚠️ that is “there was nothing to select from”, not “nothing was selected”.

Where the cut lands, the body ends with ... [truncated by token budget]. A link that was cut through the middle is removed rather than handed to you half-written.

⚠️ The budget does not lower the bill. You are charged for the pages we fetched, which is what cost us the browser — see max_pages above. Setting max_tokens buys you back context, not credits.

🚨 depth changes how much you get for the same one credit, in both directions. Measured on 2026-09-17: a 4gamer article is 8,654 estimated tokens on the default path and 1,587 with depth:"deep" — the deep read strips the furniture. The Hacker News front page is 2,430 by default and 14,636 deep, because a deep read harvests every link. Same URL, same price, and nothing tells you which you are about to get. Read rawTokens and decide from there.

⚠️ Do not set it when you want the whole page. Building a catalog from a reference table is the clearest case: a budget silently gives you a shorter page, and a shorter page gives you a shorter catalog that still looks complete.