Skip to main content
Sometimes an agent does not want ten blue links — it wants the answer, or it wants the relevant passages already cut to fit its context window. Through AgentKey your agent gets both with the same three calls — find_tools, describe_tool, execute_tool — from Perplexity’s search with page contents, Brave’s LLM-context chunks and summarizer, and Tavily’s generated answers, URL extraction, site crawl and site map.

​
What you can ask

  • “Answer with citations: what changed in Python 3.14’s free-threading support?”
  • “Give me about 8,000 tokens of grounded web context on the EU Digital Markets Act, from the last month.”
  • “Pull the page contents for the top results on ‘Rust async cancellation safety’, capped at 4,000 tokens per page.”
  • “Extract these three URLs as markdown and rank the passages by relevance to ‘pricing changes’.”
  • “Only use docs.python.org and peps.python.org as sources.”
  • “Summarize what the web says about this startup, with the source URLs.”
  • “Map the structure of docs.example.com, then crawl the API section.”

​
What’s covered

Direct answers
  • A grounded answer to a question with optional citations and entity metadata, in a chosen country backend and language, from Brave
  • A quick or more detailed generated answer returned alongside Tavily search results
  • A summary of the web results for a query from Brave’s summarizer, requested as part of a web search
Context shaped for a model
  • Brave LLM context: web-content chunks with budgets for the number of URLs, snippets per URL, tokens per URL and total tokens; a strict, lenient or balanced relevance threshold; freshness, country and language; optional source metadata and local results
  • Perplexity search with relevant page contents: result count, total and per-page token limits, country, language filters, a domain allow-list, recency from the past hour to the past year, and publish or last-updated date windows
  • Tavily search with each page’s cleaned content as markdown or text, and relevance-ranked chunks per source at advanced depth
Content from URLs you already have
  • Tavily extract: one or more URLs returned as markdown or plain text, at basic or advanced depth for tables and embedded content, optionally with images and with passages re-ranked against a query
  • Tavily crawl: traverse a site from a base URL, exploring many paths in parallel with built-in extraction
  • Tavily map: the structure of a site from a base URL

​
How to use it

Pass the request in plain language and let the router pick the tool:
Always describe_tool before executing — Tavily’s answer rides on its search call, Brave’s LLM context takes a query with token budgets, Perplexity takes max_tokens and max_tokens_per_page, and per-call costs differ.
Building with code? The same capability is available as a fixed-contract Scenario Method with automatic provider failover: web.answer answers a question from live web search results and returns the sources the answer was grounded in. For a single page’s text use web.scrape.

​
FAQ

Yes — every answer tool here searches the live web first and returns the sources it used, so your agent can cite or verify them. Treat the answer text as a draft and the sources as the evidence.
Fetched at request time. Brave and Perplexity accept freshness or recency filters; Perplexity also filters by publish and last-updated dates.
Brave’s LLM context takes a country and a language; Perplexity takes a country and ISO 639-1 language codes; Tavily is scoped by domain lists. Questions can be asked in any language.