Tavily for AI agents

Tavily was built for the retrieval step of a language-model pipeline, and it shows in what a call returns: not just links, but a relevance score per result, the full page text if you ask for it, and, if you ask, a written answer assembled from the results. For an agent that wants to read once and move on, that is often the whole job in one cent.

What Tavily does well

An answer in the same call. include_answer: true returns a synthesized answer alongside the ranked results. For a question with a factual answer, the agent can quote it and cite the results without a further step.

Relevance scores. Every hit carries a score, which lets a RAG pipeline threshold rather than take the top N blindly.

Content per hit. include_raw_content returns the full page text for each result, as markdown if you prefer, which removes the scrape step for pages that render server-side.

Topic and time filters. A topic of general, news, or finance, a time_range of day, week, month, or year, and include or exclude domain lists.

Extraction with JavaScript. tavily/extract renders JavaScript and returns clean content for up to five URLs in one call, which is the middle ground between Exa's cheap no-JS contents and a full headless browser.

Actions and prices through Vaaya

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

A real call

{
  "service": "tavily",
  "action": "search",
  "params": {
    "query": "what changed in Stripe's Machine Payments Protocol at Sessions 2026",
    "topic": "news",
    "time_range": "month",
    "include_answer": true,
    "max_results": 8
  },
  "max_cost_cents": 1
}

Eight scored results and a written answer for a cent. A ceiling of zero is refused before Tavily is contacted.

When to pick Tavily, and when not to

Pick Tavily when the deliverable is an answer rather than a reading list, when relevance scores feed a threshold, or when the query is news or finance and the topic filter helps. Pick the extractor when a page needs JavaScript and five URLs at a time is enough.

Pick Exa for descriptive queries where recall matters most, for people and company discovery, and for sweeps beyond twenty results. Pick Brave as an independent index for corroboration. Pick Perplexity when the freshness of each page decides the answer. Pick Linkup when you want a cited answer with an option for a structured JSON output.

Gotchas

Answer synthesis is only as good as the top results; for official-documentation queries, corroborate with Exa before trusting the answer. The extractor takes at most five URLs per call; batch accordingly. Raw content increases response size substantially; request it only when the agent will read it.

Tavily's documentation is at tavily.com. The search recipe is at /recipes/search-scrape.

Questions

What is the Tavily API good for?

Search tuned for retrieval-augmented generation: results carry relevance scores, the agent can ask for a synthesized answer in the same call, and raw page content can be returned per hit as markdown. It also extracts clean content from up to five URLs per call, with JavaScript rendered.

How much does Tavily cost through Vaaya?

One cent per search, including the optional answer and raw content. One cent per extract call for up to five URLs. Priced before the call and billed only on success.

Tavily or Exa for an agent?

Exa for recall on descriptive queries and for sweeps of up to a hundred results. Tavily when you want the answer written for you, when the results need relevance scores for a RAG pipeline, or when a topic filter for news or finance matters.