Scrape one URL

Strip navigation, ads, and boilerplate; keep main article content only

null = auto-detect, true = force JS rendering, false = skip JS rendering

Milliseconds to wait after JS rendering before extracting content

Return per-field evidence alongside the 'json' format: each top-level scalar schema property comes back as a Basis with an honest status. Requires a JSON schema; off by default, and when off the request sent to the model is unchanged.

CSS selector to narrow content extraction to a specific element

XPath expression to narrow content extraction to a specific element

Query string for BM25/cosine chunk filtering (used with chunkStrategy)

Filtering algorithm to rank chunks against query

Number of top chunks to return after filtering (default: 5)

Per-request proxy URL (http/https/socks5). Overrides global config.

Rotation strategy for proxyList (default: sticky_per_host)

2-letter ISO 3166-1 alpha-2 country code for residential proxy egress (e.g. 'us', 'gb')

Override stealth mode for this request (null = use global config)

BYOK: LLM provider API key for structured extraction (overrides server key)

Per-request LLM provider override

Per-request LLM model identifier override

Per-request LLM base URL override for Chat Completions-compatible or Responses-compatible providers (e.g. https://api.deepseek.com/v1)

User-supplied instructions appended to the summary system prompt (capped at 500 chars)

Maximum bytes of scraped content sent to LLM for 'summary' format (capped at 200 KB)

Pin this request to a specific renderer. A non-auto browser tier implies renderJs:true unless renderJs:false is set. impersonated-http is a JS-less tier that presents a real Chrome TLS fingerprint; it never implies renderJs, and combining it with renderJs:true is rejected.

End-to-end deadline budget in milliseconds (default: 8000, max: 60000)

When true, response includes debugExtraction trace with every candidate the extractor considered

Plain-language monitor goal for the meaningful-change judge

Run the LLM meaningful-change judge on a changed page (requires goal)

Capture the whole page rather than the viewport, for formats: ["screenshot"].

How old a cached copy of this fetch may be, in milliseconds. 0 always fetches. Clamped to a 24 hour ceiling. Extraction options (formats, schema, selectors) are not part of the cache key, so a second pass over the same page reuses the fetch.

Set false to read the cache without writing to it.

Top-level advisory (e.g. fallback render notice)