Brave Search for AI agents

Every semantic search engine in the network is excellent at finding pages that mean what you asked. None of them can tell you whether the page it found is the page everyone else would find, because they all rank by the same kind of signal. Brave is different in the way that matters for corroboration: it runs its own crawler and its own index, and it ranks by keywords. When an agent has one strong result and needs a second, independent opinion before acting on it, Brave is that opinion at a cent.

What Brave does well

An independent index. Brave's results come from Brave's own crawl, not from Google or Bing and not from an embedding model. For a corroboration step, that independence is the whole point: two engines that fail differently are worth more than one engine queried twice.

Keyword precision. When the query is an exact phrase, an error string, a product name, or a version number, keyword ranking beats semantic ranking, and Brave is the keyword engine in the network.

News with structure. The news endpoint returns recent articles with the source, the age, and a breaking flag, filtered by past day, week, month, or year. For "what happened this week" questions, this is the shape an agent wants.

Filters that scope a search. Country, language, freshness including a custom date range, safe search, pagination, a result-type filter, and extra snippets per hit.

Actions and prices through Vaaya

Priced before the call, billed only on success, under max_cost_cents.

A real call

{
  "service": "brave",
  "action": "news",
  "params": {
    "q": "Agentic Payments Alliance Visa Mastercard Fiserv",
    "freshness": "pw",
    "count": 10
  },
  "max_cost_cents": 1
}

Ten articles from the past week with source and age, for a cent. A ceiling of zero is refused before Brave is contacted.

When to pick Brave, and when not to

Pick Brave as the second engine whenever a result will be acted on: run Exa and Brave together, take what both agree on. Pick it first when the query is an exact string. Pick the news endpoint for anything that happened recently.

Pick Exa for descriptive queries and for content inline with the results; Brave returns snippets, not pages. Pick Tavily or Linkup when you want a written answer. Pick Perplexity when the last-updated date of each page decides the answer. Pick Valyu's academic search for papers.

Gotchas

Brave returns snippets, not full text; chain a read through Exa's contents endpoint or Tavily's extractor when the agent needs the page. Web search caps at twenty results per call. Keyword ranking is literal: a descriptive query that works on Exa will underperform here, and the reverse is also true.

Brave's documentation is at brave.com/search/api. The search recipe is at /recipes/search-scrape.

Questions

Why would an agent use Brave Search alongside Exa?

Because Brave's index is its own crawl, not a re-ranking of another engine's results. When a semantic search returns something that looks right, a Brave keyword search over an independent index is the cheapest way to corroborate it. Different index, different failure modes.

How much does Brave Search cost through Vaaya?

One cent per web search of up to twenty results, and one cent per news search of up to fifty. Priced before the call, billed only on success, under the maximum cost the agent states.

What does the Brave news endpoint add?

Recent articles with source, age, and a breaking flag, with freshness filters of past day, week, month, or year. It is the right call for current-events questions where a web search keeps returning evergreen pages.