Skip to main content
Half of what an agent does with the web is read a page it already has the address for — an article, a product page, a PDF, a changelog. Raw HTML wastes context; what the model needs is the main content as clean markdown. Through AgentKey your agent gets exactly that with the same three calls — find_tools, describe_tool, execute_tool — from Firecrawl or Jina Reader.

​
What you can ask

  • “Read this article and give me the main text as markdown, without the navigation and footer.”
  • “Scrape this PDF and return the text page by page.”
  • “Fetch this page as a visitor from Germany would see it, in German.”
  • “Extract the title, price and SKU from this product page as JSON.”
  • “Click ‘Load more’ twice, then scrape the list.”
  • “Give me the mobile version of this page and a screenshot.”
  • “Return only the links on this page.”
  • “Use a cached copy if you have one from today; otherwise fetch fresh.”

​
What’s covered

Clean content
  • Markdown, HTML, raw HTML or plain text from a URL; main content only by default, with headers, navigation and footers removed by a deterministic filter
  • An optional extra cleaning pass for residual boilerplate — cookie banners, share widgets, related-article lists — keeping headings, tables and links
  • Include or exclude HTML tags; target or remove elements by CSS selector; ad and cookie-popup blocking; base64 images stripped from the output
  • Summaries of a page’s links and images, generated alt text for images, and a token budget for the returned text
Formats and extraction
  • Several output formats in one call, including structured JSON extracted by an LLM against a schema you provide
  • Screenshots and full-page shots; mobile-device emulation and a custom viewport
Documents
  • PDFs parsed into markdown page by page, or returned whole as base64
Dynamic pages
  • Page actions performed before scraping, a fixed wait, or a wait for a CSS selector
  • A choice of fetch engine on Jina — direct, headless browser, Cloudflare browser rendering or a reader model — plus iframe and shadow-DOM content
  • Custom headers and cookies, and a named browser profile that keeps cookies and local storage between scrapes
Caching and freshness
  • Accept a cached copy newer than a given age, insist on cache only, or run in lockdown mode where the target is never contacted
  • Turn caching off for a request, or ask for zero data retention
Location and proxies
  • A country and preferred languages — the matching proxy, locale and timezone are emulated
  • Basic or enhanced proxies, or automatic escalation when the basic fetch fails; a proxy country or your own proxy URL on Jina; timeouts and a custom robots.txt user agent

​
How to use it

Pass the request in plain language and let the router pick the tool:
Always describe_tool before executing — Firecrawl’s options are body parameters, Jina’s are x-* header options (its GET form takes the target URL as the path), and Firecrawl’s cost is metered on what the page needs.
Building with code? The same capability is available as the fixed-contract web.scrape Scenario Method: one URL in, readable markdown or plain text with title and basic metadata out, with automatic failover across the scrapers above. See Scenario Methods.

​
FAQ

Yes. Firecrawl renders the page and can wait or perform actions first; Jina can be switched to a headless browser engine. If a site answers with a bot challenge instead of content, see Anti-bot pages.
A scrape returns what the page serves to the fetcher. These tools do not unlock anything; if you hold a session you are entitled to, pass its cookies or headers and the page is fetched as that session sees it.
Fetched at request time unless you ask for a cached copy. Firecrawl’s cache-age, cache-only and lockdown options, and Jina’s cache tolerance, let you trade freshness for speed explicitly.