Skip to main content
Web search is the one capability almost every agent needs and the one a base model cannot do on its own: the answer to “what is true right now” lives on pages published after training. Through AgentKey your agent searches the same way it does everything else — find_tools, describe_tool, execute_tool — and chooses between several engines behind one key: Google results through Serper, Brave’s own independent index, Tavily’s LLM-oriented search, Exa’s neural and keyword retrieval, and Parallel for running several queries in one call.

​
What you can ask

  • “Find the official documentation for the Rust tokio crate.”
  • “What has been written about AgentKey in the past week, in English?”
  • “Search only arxiv.org and github.com for retrieval-augmented generation benchmarks.”
  • “Run these five product-name queries at once and give me the top result for each.”
  • “Give me German-language results from Germany for ‘Wärmepumpe Förderung 2026’.”
  • “Find pages about a CLI tool that syncs OpenAPI specs to a docs site, even if they don’t use those words.”
  • “Page through the first three pages of Google results for this query and dedupe the domains.”
  • “Return the full page text with each result so I can read the top three sources.”

​
What’s covered

Query controls
  • Search operators — site:, exact phrases, exclusions — on the engines that support them; Brave lets you switch operator parsing off for literal queries
  • Spell correction and autocorrect, on by default and switchable
  • Query suggestions and autocomplete for a partial query
  • Custom re-ranking of Brave results with Goggles, hosted or inline
Country, language and location
  • A country code to bias results, and separate codes for the result language and the interface language
  • A named location such as “New York, United States” for Google results; latitude / longitude, city, state or postal-code hints for Brave
Freshness
  • Past hour, day, week, month or year on Google results
  • The last 24 hours, 7 days, 31 days or 365 days, or a custom date range, on Brave
  • A relative time range or explicit start and end dates on Tavily
Pagination and volume
  • Page numbers on Google results, an offset on Brave, and a per-call result count on every engine
  • Several queries in a single Parallel call, each with its own ranked results
Result shapes
  • Ranked organic results with title, URL and snippet; extra snippets on Brave
  • Brave result filters to keep only certain blocks — discussions, FAQ, infobox, news, videos, locations — and an optional summary
  • Tavily options to include a generated answer, each page’s cleaned content as markdown or text, and images; to restrict or exclude domains; topic modes for general, news and finance; basic, fast or advanced depth
  • Exa retrieval modes — auto, neural, keyword, fast — with full page text or highlighted snippets per result

​
How to use it

Pass the request in plain language and let the router pick the engine:
Always describe_tool before executing — the same idea has different parameter names on each engine (gl / hl on Google results, country / search_lang on Brave, tbs versus freshness versus time_range), and per-call costs differ.
Building with code? The same capability is available as a fixed-contract Scenario Method with automatic engine failover: web.search takes a query with optional country, language, time range, domain include / exclude lists and a pagination cursor, and returns a normalized result page.

​
FAQ

Google results through Serper take any Google country and interface language code, plus a free-text location. Brave takes a fixed list of countries and result languages covering the Americas, Europe and most of Asia-Pacific, including simplified and traditional Chinese. Tavily’s country boost applies to its general topic. Queries themselves can be in any language.
Every call goes to the engine at request time; nothing is served from an AgentKey cache. Freshness filters narrow results by the engine’s own discovery or publication date.
Google results when you want familiar ranking and operators; Brave when you want an independent index or result-type filters; Exa when the query is a description rather than keywords; Tavily when you want cleaned page content or an answer alongside results; Parallel when you have many queries to run at once. execute_tool calls exactly the engine you name — switch engines yourself if one fails, or use web.search, which fails over for you.