Exa

Point an agent at a conventional search API and you get what a person would get: ten blue links, ranked for clicks, pointing at pages full of navigation, cookie banners, and advertising. The agent then fetches each one, strips the HTML badly, and spends most of its context on furniture.

Exa is built for the other reader. Its index is searched by meaning rather than keywords, and every result can come back with the page’s clean extracted text, so the agent gets content instead of markup.

Search and read in one call link

  pip install exa-py
  

  from exa_py import Exa
exa = Exa(api_key="...")

results = exa.search_and_contents(
    "engineering blog posts about migrating off Kubernetes",
    num_results=5,
    text=True,
    start_published_date="2026-01-01",
)

for r in results.results:
    print(r.title, r.url)
    print(r.text[:500])
  

Describing what you want in a sentence works better here than keyword stuffing, because the query is embedded rather than matched. The filters matter as much as the query: publication date, domain include and exclude lists, and result type all narrow the search before ranking, which is how you keep an agent away from SEO farms.

Two endpoints beyond plain search are worth knowing:

• answer — a direct answer with the sources it used. Good when you want one fact and a citation rather than a reading list.
• research — a multi-step agentic search that plans, gathers, and returns structured output with web-grounded citations. Slower and more expensive per call; appropriate when the question genuinely needs several searches.

There’s an official MCP server, so an MCP-capable agent can use all of this without you writing a client at all.

Costs, and how not to be surprised link

Pricing is per thousand requests and differs sharply by endpoint — plain search is cheap, the research endpoint costs several times more, and extra results and page summaries bill on top. New accounts get free credits with a monthly top-up, which is enough to build and evaluate against.

Three habits keep the bill sane, and they also make the agent better:

• Cache aggressively. The same query from ten users should hit the API once — Upstash with a day’s expiry is the usual answer.
• Default to search. Reach for research only when a single search genuinely can’t answer the question. It’s easy to wire the expensive endpoint in as the default and discover the cost later.
• Ask for fewer results and less text. More context is not more accuracy; five clean results usually beat twenty noisy ones.

Where it fits among the alternatives link

The distinction that matters: Exa searches the public web, so it answers “what’s out there about this”. It has nothing to say about your internal documents — for those you build your own index, often from documents parsed with Docling or MinerU.

And treat every retrieved page as untrusted input. Content an agent fetches from the web can contain instructions aimed at the agent; that it came back from a search does not make it safe to follow.

Next link

Changes made with AI need version control around them → Git & Collaboration