First scrape

Read the result

After a scrape finishes, the telemetry bar under the control strip tells you what happened:

• The final URL — where the page actually landed, after redirects.
• selector cache hit — a prior run already learned CSS selectors for this site, so extraction made no LLM call.
• N tokens · $X.XXXX — what any LLM step cost, priced from the provider’s table.
• Nms — how long the fetch took.
• N chars — the size of the Markdown.

The result itself opens below as rendered Markdown. The screen keeps it while you navigate away and back — nothing is lost to a trip to History.

Formats

The format chips pick what one fetch produces: markdown, llm (token-lean text for LLM input), text, json, html (cleaned) and raw (as served). Each picked format gets its own tab in the result, all from the same fetch — picking json does not refetch the page. Markdown always renders, whatever you picked.

The Viewport tab is not a format: it renders the page in a browser and screenshots it, on demand. Set the size first in Advanced (Viewport, WxH).

Extractors

The chip row after the formats picks one extractor for the run:

• off — plain cleaned Markdown.
• auto — dispatch by URL pattern to a vertical if one matches.
• schema — describe what you want and let an LLM draft a JSON Schema, or pick one you saved. See Schemas.
• a vertical — one of the 27 zero-LLM extractors (GitHub, Hacker News, YouTube, arXiv, PyPI, npm, and the rest). Hover a chip for what it matches.

A vertical costs nothing and needs no key. A schema costs LLM tokens on the first runs, then drops to zero once selectors are learned — see Extract.

Keep the result

Above the tabs: Copy and Save as file always act on the visible tab, under the tab’s own name (page.md, page.json, page.raw.html…). The Viewport tab saves the screenshot as page.png. Discard result clears the pane.

Summarize condenses the page to 1–20 sentences without refetching. Brand pulls the page’s palette, fonts, logo and favicon at zero LLM cost. Copy as… shows the exact run as a CLI command, an MCP tool call, or a Go snippet — useful for turning a click into a script.

Everything is also in History, with the tokens and cost per run.

When it fails

“This site blocked us” — the site served a bot challenge. Press Retry with Browser render; if it still fails, the site’s defenses hold. Magpie reports challenges; it does not defeat them.

“Extraction failed — the fetched page is below” — the fetch worked, the LLM step did not. Usually a missing key (an Open Settings button appears) or a spent cost cap. Set the key in Keys and press Re-extract on a format tab.

Empty Markdown on a JavaScript-heavy site — the static fetch returned a shell. Switch Render to Browser.

The data you want is in the network tab, not the page — open Advanced, set Capture XHR to a pattern like /api/.*, re-run with Browser render, read the XHR tab.

More symptoms live in Troubleshooting.