Exa is a web search API built for AI agents, not for people. An agent sends a query and gets back ranked results with the page text or short highlights: the passages of each page that answer the query, ready to drop into a prompt.
The difference from a search wrapper: Exa runs its own index and its own search models. It doesn't resell Google or Bing results. It's the search behind Cognition's Devin, among others, and claims 500,000 developers.
What it does
| Product | What it does | Price |
|---|---|---|
| Search | Web search with text or highlights, in about 450 ms to 1 s | From $4 per 1,000 requests |
| Deep Search | Multi-step search with reasoning, structured output and citations, 4 to 15 s | $12 to $15 per 1,000 |
| Contents | Clean page text, highlights or summaries for known URLs, PDFs included | $1 per 1,000 pages |
| Answer | An LLM answer to a question, with citations | $5 per 1,000 requests |
| Monitors | Scheduled searches that report new events on the web | $2 per 1,000 requests |
| Agent | Async research, list building and enrichment, with structured output | $0.012 to $1.00 per run |
Categories narrow a search to a vertical: news, companies, people, research papers and financial data. The Exa Library adds structured collections searchable with the same call: companies, people, job postings, SEC filings, earnings calls, patents, court opinions, podcasts and more.
Exa Connect plugs paid data providers into Agent runs (contact data, private markets, web analytics), billed per provider call.
For business people
The problem it solves. An AI assistant without search answers from stale training data. Classic search APIs return 10 blue links built for humans, and the agent then scrapes each page. Exa returns the relevant passages directly, which cuts the tokens sent to the model. Exa claims up to 90% fewer tokens with highlights.
Where it fits:
- Grounding a chatbot or voice agent in current facts.
- Coding agents that need current docs and code.
- GTM and enrichment: find every company or person matching a description, then enrich them with structured fields. HubSpot and monday.com are named customers here.
- Monitoring: a daily search for news on a list of accounts.
Pricing (November 2026). Pay as you go, no subscription:
| Plan | What you get |
|---|---|
| Free | $10 of credits every month, plus a $10 onboarding bonus, no card |
| Developer | Pay per request, up to 25 searches per second, SOC 2 Type II |
| Enterprise | Zero data retention, HIPAA, SSO, SLAs, custom indexes, up to 1,000 results per search |
The free $10 covers up to 2,500 Lite searches a month. A base request includes up to 10 results. Each extra result costs $1 per 1,000.
Trust and compliance: SOC 2 Type II, GDPR and CCPA, HIPAA with a BAA. Zero data retention, where queries are never stored or trained on, is part of Enterprise, so check the terms before sending sensitive queries on the self-serve plans.
The company. Exa Labs started as Metaphor, founded by Will Bryk and Jeff Wang, and renamed itself Exa in 2024. It raised a Series B in September 2025 at a reported $700 million valuation. A May 2026 Series C led by Andreessen Horowitz, at a reported $2.2 billion, comes from press reports, not a company announcement I could verify.
For technical people
How it searches. Exa trains its own embedding and retrieval models and serves them from its own vector database: 100 billion documents served, 2 trillion URLs tracked, and about 1.9 billion URLs crawled a day, by its own count. Retrieval works on meaning, not only keywords, so a description of the page you want often works better than keywords. Exa 3.0 (October 2026) added the Lite tier.
Search types trade latency for depth:
type |
Latency | Price per 1,000 (up to 10 results) |
|---|---|---|
lite |
~450 ms | $4 |
auto |
~1 s | $7 |
deep-lite |
~4 s | $12 |
deep |
4 to 15 s | $12 |
Auth: x-api-key or Authorization: Bearer. The official SDKs are exa-py and exa-js, and both read EXA_API_KEY from the environment.
from exa_py import Exa
exa = Exa() # reads EXA_API_KEY
result = exa.search(
"independent watch brands making field watches in Germany",
type="auto",
contents={"highlights": True},
)
for r in result.results:
print(r.title, r.url)
curl -X POST https://api.exa.ai/search \
-H "x-api-key: $EXA_API_KEY" \
-H "Content-Type: application/json" \
-d '{"query": "latest open-weight LLM releases", "type": "auto", "contents": {"highlights": true}}'
Useful parameters: includeDomains and excludeDomains, date filters, category, the number of results, and livecrawl policies on Contents to force a fresh fetch. Search can stream results as server-sent events.
Agent interfaces:
- A hosted MCP server at
https://mcp.exa.ai/mcp, usable without a key at low volume, with OAuth or a key above that. Tools cover web search, code search, contents and research. - Agent skills:
npx skills add exa-labs/agent-skills. - Docs as Markdown: append
.mdto any docs URL, or start fromexa.ai/docs/llms.txt. The OpenAPI spec is the source of truth for schemas. /searchand/contentsalso accept x402 pay-per-request payments, so an agent can pay without an account.
Limits: 10 searches per second on the free plan, 25 on Developer, 50 concurrent Agent runs. Contents handles JavaScript pages and PDFs, but it isn't a site crawler: for a full site, use a crawler.
Limits and caveats
- The benchmarks are Exa's own. The home page shows Exa ahead of Perplexity, Brave and Parallel on accuracy. Exa publishes its evals, and some harnesses are open source, but treat vendor charts as claims.
- Semantic search can miss exact matches. For an exact phrase, a part number or a quoted name, a keyword engine can do better.
- Costs add up past 10 results and with summaries, Deep Search and Agent runs. Metered Agent runs cap at $5 (
auto) or $20 (ultra) by default, so set the cap. - Zero data retention is an Enterprise feature.
Alternatives
| API | Approach |
|---|---|
| Tavily | Search with cleaned snippets, simple for prototypes |
| Perplexity Sonar | Returns a written answer with citations |
| Brave Search API | Independent keyword index, links first, cheap |
| Parallel | Search and deep research for agents |
| Firecrawl | Scraping and crawling, with a search endpoint |
Firecrawl is the complement, not a replacement: Exa finds the pages, Firecrawl crawls and extracts whole sites. My notes on it: see Firecrawl.
My take
Exa is the search layer I'd pick for agent work that needs to find things by description: "companies like X", "people who wrote about Y", "every new open-weight model this month". The highlights are the real gain: fewer tokens and less scraping glue.
My stack today uses Perplexity for research answers and Firecrawl for page content. Exa would sit between them for list building and monitoring, where a description beats keywords. The free $10 a month is enough to test that on real prospecting queries before paying anything.
Further reading
The documentation, with Markdown versions for agents:
Every price, endpoint by endpoint:
The MCP server for Claude, Cursor and Codex:
The published benchmarks: