---
title: "Exa AI API guide (2026): search types, pricing, API key and MCP"
category: "guides"
url: https://keirolabs.cloud/blog/exa-ai-api-guide-2026
---

## The 60-second version

**Exa is a search engine built for AI: instead of matching keywords, it uses embeddings to find pages that mean what your query means, and it can return their full text in the same call.** Developers use it to ground agents, find similar companies or papers, and pull clean page content for RAG.

- **Get a key:** sign up at [dashboard.exa.ai](https://dashboard.exa.ai). New accounts get **$10 in credit that resets each month**, no card required, plus a one-time $10 for finishing onboarding.
- **Price:** search is **$7 per 1,000** on the default `auto` and `fast` types, **$4** on `instant`, $12 to $15 on the `deep` types, for up to 10 results. Page content is $1 per 1,000 pages per content type.
- **Endpoints:** `/search`, `/contents`, `/answer`, plus Agent runs, Monitors and Websets.
- **MCP:** a hosted server at `https://mcp.exa.ai/mcp` works **without an API key**, rate-limited.

**For most agents: `type: "auto"` with `highlights` instead of full `text`.** Highlights give the model the relevant passages for far fewer tokens.

## What is Exa AI?

**Exa is a neural web search API: it indexes the web with embeddings and returns the pages closest in meaning to your query, along with their content if you ask.** The company behind it was called Metaphor until 2024.

That changes what a good query looks like. A keyword engine wants "best vector database"; Exa does better with a description of the page you want, such as "a blog post comparing vector databases on recall and cost, written by an engineer who ran benchmarks". It also offers what keyword engines cannot: give it a URL and it finds pages like that one, which is how people use it for lead lists, competitor research and literature search.

| | Keyword search API (Brave, SerpAPI) | Exa |
|---|---|---|
| Matches on | Words in the page | Meaning of the page |
| Best query | Short keywords | A description of the ideal page |
| Find pages similar to a URL | No | Yes |
| Page content in the same call | Usually no | Yes: text, highlights or summary |
| Categories | Web, news, images | Company, people, news, research papers, financial reports and more |

We test it against ten other APIs in [best web search API for AI agents](/blog/web-search-apis-for-agents-2026).

## How do you get an Exa API key?

1. **Sign up** at [dashboard.exa.ai](https://dashboard.exa.ai).
2. **Create a key** under API Keys and copy it.
3. **Store it as `EXA_API_KEY`.** Both official SDKs read that variable automatically.
4. **Use the free credit to test.** $10 covers about 1,400 default searches or 2,500 instant ones, and the free balance resets to $10 on the first of each month ([Exa pricing](https://exa.ai/pricing)).

Pay as you go has no subscription and no minimum. Enterprise adds volume discounts, invoicing, custom rate limits and zero data retention.

## How do you make your first Exa search?

**Install `exa-py` (or `exa-js`), create a client and call `search` with the content you want back.**

```python
# pip install exa-py
from exa_py import Exa

exa = Exa()  # reads EXA_API_KEY

result = exa.search(
    "engineering blog posts benchmarking vector databases on recall and cost",
    type="auto",
    num_results=5,
    contents={"highlights": True},   # relevant passages, not whole pages
)
for r in result.results:
    print(r.title, r.url)
    print(r.highlights[:1])
```

The same request over HTTP:

```bash
curl -X POST https://api.exa.ai/search \
  -H "x-api-key: $EXA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"query": "startups building search APIs for AI agents", "type": "auto",
       "numResults": 5, "contents": {"highlights": true}}'
```

Each result can include `title`, `url`, `publishedDate`, `author`, `text`, `highlights`, `summary` and `favicon`. The response also returns **`costDollars`**, an estimate of what the call cost, which makes it easy to track spend per request ([Search reference](https://exa.ai/docs/reference/search)).

## Which Exa search type should you use?

| Type | Price per 1,000 (up to 10 results) | Use it for |
|---|---|---|
| `instant` | $4 | Lowest latency: autocomplete, high-volume lookups |
| `fast` | $7 | Quick agent tool calls |
| `auto` (default) | $7 | Most queries; Exa picks the approach |
| `deep-lite` | $12 | Harder queries that need more searching |
| `deep` | $12 | Multi-step research questions |
| `deep-reasoning` | $15 | Questions that need reasoning over many sources |

**Start with `auto`.** Drop to `instant` when speed matters more than recall, and reach for the deep types only when `auto` misses. The deep types also have a lower rate limit: 5 queries per second instead of 10.

Other parameters worth knowing:

- **`numResults`**: 1 to 100, default 10. Results above 10 cost $1 per 1,000 extra results.
- **`category`**: `company`, `people`, `news`, `publication`, `personal site`, `financial report`. The company and people categories do not accept date filters or excluded domains; sending them returns a 400.
- **`includeDomains` / `excludeDomains`**: up to 1,200 entries each, with wildcard subdomains.
- **`startPublishedDate` / `endPublishedDate`**: ISO 8601 dates.
- **`maxAgeHours`**: controls freshness of page content. `0` always fetches live, `-1` always uses cache. It replaces the old `livecrawl` option, which is deprecated.

## How much does the Exa API cost?

**Exa bills searches per request and page content per page, per content type.** Rates from [exa.ai/pricing](https://exa.ai/pricing), checked October 15, 2026:

| Item | Price |
|---|---|
| Search, `instant` | $4 per 1,000 requests |
| Search, `auto` / `fast` | $7 per 1,000 requests |
| Search, `deep-lite` / `deep` | $12 per 1,000 requests |
| Search, `deep-reasoning` | $15 per 1,000 requests |
| Each result above 10 | $1 per 1,000 results |
| Contents: text, highlights or summary | $1 per 1,000 pages, per content type |
| Answer | $5 per 1,000 requests |
| Monitors | $15 per 1,000 requests |
| Agent runs | $0.012 (minimal) to $1.00 (xhigh) per request |

Two things make real bills higher than the headline. **Content types stack:** asking for text and highlights on the same page counts as two. **And content is per page:** ten results with highlights add ten pages to the bill. So a default search with highlights on 10 results is about **$0.017**, not $0.007. Our [Exa pricing breakdown](/blog/exa-api-pricing-2026) works through bills at 10,000 to a million searches.

## What are Exa's other endpoints?

- **`/contents`** takes URLs and returns their text, highlights or summary. Use it when you already have links; it allows 100 requests per second.
- **`/answer`** searches and writes a cited answer, at $5 per 1,000. It also has an OpenAI-compatible chat completions interface.
- **Agent runs** carry out longer research tasks, priced by effort level, with search calls inside them at $0.005 each.
- **Monitors** re-run a search on a schedule and return what changed, at $15 per 1,000.
- **Websets** build lists of entities (companies, people) that match criteria, then enrich them.

## How do you use Exa with MCP and Claude Code?

**Exa's hosted MCP server is at `https://mcp.exa.ai/mcp` and needs no API key to start:** keyless use is free and rate-limited ([Exa MCP docs](https://exa.ai/docs/reference/exa-mcp)). Add your key or sign in with OAuth when you hit 429s.

In Claude Code, install the official plugin, or add the server directly:

```bash
claude plugin install exa@claude-plugins-official
# or
claude mcp add --transport http exa https://mcp.exa.ai/mcp
```

The server exposes `web_search_exa` and `web_fetch_exa` by default, `agent_run` with a key or OAuth, and `web_search_advanced_exa` if you opt in. See how it compares with nine other servers in [the best MCP servers for web search](/blog/best-mcp-servers-web-search-scraping-2026).

## What are the Exa API rate limits?

| Endpoint | Default limit |
|---|---|
| `/search`, `/answer`, chat completions | 10 queries per second |
| `/search` with `deep-lite`, `deep`, `deep-reasoning` | 5 queries per second |
| `/contents` | 100 queries per second |
| Agent runs | 5 per second, 50 active runs |
| Websets | 20 per second |

Limits can vary by plan; Enterprise sets custom ones. **10 queries per second is 600 a minute**, enough for most agents but low for bulk enrichment jobs, where `/contents` at 100 per second is the better fit. Other providers' limits are in [search API rate limits](/blog/search-api-rate-limits-2026).

## What are the best Exa alternatives?

| Alternative | Price per 1,000 searches | Choose it when |
|---|---|---|
| **Keirolabs** | from $0.25 (search), $0.10 (Google SERP) | You want search plus page content at a lower price, 1,250 free credits a month |
| **Tavily** | $8 pay as you go, $5 on Growth | You want the most framework integrations |
| **Brave Search API** | $5 | You want a large independent keyword index |
| **Parallel** | from $1 | You want dense excerpts at the lowest per-call price |
| **Perplexity Search API** | $5 | You want Perplexity's ranking |

Prices from each vendor's site or our dated breakdowns: [Tavily](/blog/tavily-api-guide-2026), [Brave](/blog/brave-search-api-pricing-2026), [Parallel](/blog/parallel-search-api-pricing-2026), [Keirolabs](https://keirolabs.cloud/pricing). Disclosure: Keirolabs is ours. Exa's find-similar and people and company search have no close equivalent in this list; if you need those, Exa is the pick. More options in [Exa alternatives](/blog/exa-alternatives-2026) and the head-to-head [Exa vs Tavily](/blog/exa-vs-tavily).

## Common Exa mistakes

- **Writing keyword queries.** Describe the page you want; Exa is built for that.
- **Asking for full `text` on every result.** Use `highlights` unless the model needs the whole page; it is cheaper in tokens, and requesting both bills twice.
- **Raising `numResults` without budgeting.** Every result above 10 costs extra, and content is billed per page.
- **Using `livecrawl`.** It is deprecated; use `maxAgeHours`.
- **Date filters on company or people searches.** Those categories reject them with a 400.

## The short answer

Exa is the search API to use when meaning matters more than keywords: semantic queries, finding similar pages, and company or people search. It costs $7 per 1,000 default searches, $4 on `instant`, plus $1 per 1,000 pages of content, with $10 of free credit every month and a free hosted MCP server. If you mainly need fast, cheap keyword results with page content, compare the alternatives above first.