Exa
The world's data for AI agents
About Exa
Exa is a search engine built for AI agents rather than for people. A traditional search API returns a ranked list of links designed for a human to scan, which is close to useless when the consumer is a language model that needs the actual page text, inside a token budget, immediately. Exa indexes the web and can return both the results and the cleaned contents of those pages in a single call, so an agent gets from a question to grounded text without you running a crawler, a parser and a cache of your own.
The problem it solves is the one every retrieval application hits about a week in. Doing this yourself means writing a fetcher, handling JavaScript rendering, respecting robots.txt, stripping navigation and boilerplate, managing rate limits, and then keeping all of that working as sites change. Exa does that part and exposes it as a small set of HTTP endpoints with SDKs and an MCP server on top. Retrieval is semantic rather than keyword based, so a query can describe the kind of page you want instead of guessing the words printed on it, which is what makes it fit agent workloads where the model knows its intent but not the vocabulary.
The company describes the product as the world's data for AI agents, covering the public web alongside private sources a customer connects. It publishes benchmark results on FRAMES, Tip-of-Tongue and Seal0, and quotes sub-180ms responses for its Instant tier. Those are Exa's own figures rather than independent measurements, so treat them as claims. Exa says roughly 500,000 developers build on it, and names Cognition, HubSpot, Firefox, monday.com, CodeRabbit, StackAI and 11x among its customers.
You create an account, load credits, and call the API with a key. There is no subscription to pick and no seat count to estimate, which matters because agent traffic is bursty in a way that seat licences model badly.
A typical loop is two calls or one. Send a query to the Search endpoint and get back matching URLs with metadata. Either follow up with the Contents endpoint to pull the cleaned text, or ask for contents inline with the search so it comes back in the same round trip. Contents supports highlights and summaries, so you can request the relevant passage rather than the whole document and keep the prompt small.
For higher level work there are endpoints that do more of the job. Answer returns a written answer with citations instead of raw results. Agent runs a multi-step research task at a selectable effort level, trading cost against depth. Monitors re-run a query on a schedule and surface what changed, which is the piece most people otherwise build badly themselves. Websets turns a query into a structured collection of entities rather than a page list.
- •Semantic search over a full web index - queries match on meaning rather than keywords, so you can ask for the type of page you want. Exa says this covers the public web plus private sources a customer connects.
- •Contents in the same request - cleaned page text, highlights and summaries returned alongside results, removing the crawl and parse step from your own stack and keeping token usage down.
- •Answer endpoint - a written, cited answer for cases where you want the conclusion rather than the sources, billed per request at a flat rate.
- •Agent endpoint with effort levels - multi-step research from minimal to xhigh effort, priced from $0.012 to $1.00 per request, so cost scales with how hard the question actually is.
- •Monitors - scheduled re-runs of a query that report changes over time, useful for competitive tracking and alerting without writing your own scheduler and diff.
- •MCP server and SDKs - a Model Context Protocol server plus official client libraries, so Claude, Cursor and other MCP-aware tools can call it without custom glue.
Developers building retrieval-augmented applications are the core audience, particularly anyone who has already tried scraping the open web and discovered how much of the work is maintenance rather than code. If your agent needs current information and you are maintaining a crawler to get it, this replaces that crawler.
Teams building research and monitoring products get the most out of the higher level endpoints. Monitors and Websets exist for competitive intelligence, lead sourcing and market tracking, which are the products people were building on top of raw search anyway.
It fits solo developers and small teams well because of the pricing shape. There is no minimum spend and no subscription, the free Starter tier renews $10 of credits monthly with no payment method on file, and a new account starts with $20, so evaluating it properly costs nothing. Startup and education projects can apply for up to $1,000 in credits on top of that. Larger engineering teams will care more about the latency claims and the private data connectors.
It is not a consumer product. There is no chat interface to sit in front of, and if you want an AI search engine to use rather than to build on, this is the wrong layer of the stack.
Pricing
Usage based, no flat monthly plan
- Starter$0/mo
- UsageNot listed
- EnterpriseContact sales
From the vendor pricing page, 2026-09-22
