E
Official Search & Web

Exa

Neural search engine for retrieving high-quality web content.

Works with: Claude DesktopClaude CodeCursor
Quick install
npx -y exa-mcp-server

How to install the Exa MCP server

Add this to your Claude Desktop MCP configuration:

{
  "mcpServers": {
    "exa": {
      "command": "npx",
      "args": [
        "-y",
        "exa-mcp-server"
      ]
    }
  }
}

Add this to your Claude Code MCP configuration:

npx -y exa-mcp-server

Add this to your Cursor MCP configuration:

{
  "mcpServers": {
    "exa": {
      "command": "npx",
      "args": [
        "-y",
        "exa-mcp-server"
      ]
    }
  }
}

The Exa MCP server gives Claude access to Exa’s neural search engine. Where Google and Brave rank pages by keywords plus authority, Exa ranks by semantic similarity. The result: better-quality sources for research questions, especially when you’re looking for long-form content.

Exa is a small but very capable team. The server tracks their API closely and is the same backend that powers their direct chat product.

Why use it

For “find me sources” workflows, neural search wins on quality. Asking Brave for “essays on the future of marketplaces” returns SEO-optimised listicles. Asking Exa returns the actual essays. The difference compounds for research-heavy work.

For solo founders writing content, doing competitor research, or studying a market, Exa’s results-by-meaning are usually closer to what you actually want than keyword results.

What it actually does

A neural search primitive: send a natural language query, get back a ranked list of pages with titles, URLs, summaries, and an Exa-generated relevance score. Optional filters: domain include/exclude, date range, content type (article, paper, video). Some servers also support similarity-based queries (find pages similar to this URL).

Practical patterns:

  • “Find serious essays about platform pricing strategy.”
  • “What are the highest-quality posts about pre-PMF marketing?”
  • “Find pages similar to this Stripe blog post about pricing power.”

Gotchas

Slower than keyword search. Exa’s neural ranker takes a few seconds per query. For “tell me what’s happening right now” lookups, Brave or DuckDuckGo are faster. For research, the delay is fine.

Exa indexes a curated subset of the web. Some smaller sites are missing. If you’re searching for content on a specific obscure domain, Exa may return nothing while Brave returns a result.

The natural pairing is Exa for discovery plus Fetch for retrieval. Exa surfaces high-quality URLs, Fetch reads them in full, Claude synthesises. For more detailed scraping (especially of JS-heavy sites), swap Fetch for Firecrawl.

If you want both broad and deep coverage, install Exa alongside Brave Search and Tavily. Claude will pick the right one based on what your prompt asks for. Three search engines covering different shapes of intent.

Exa MCP server: FAQs

Is the Exa server official?

Yes. Exa Labs ships and maintains exa-mcp-server. It's the same API that powers their direct chat product, exposed as MCP.

How is neural search different from Google or Brave?

Traditional search ranks pages by keyword match plus authority signals. Exa ranks by semantic similarity to your query and was trained specifically to find high-quality long-form content. For research questions ('find me serious essays about X'), it tends to surface better sources than keyword search.

What does it cost?

Exa has a free tier with rate limits and several paid tiers. Pricing is per request, scaling with volume. For most personal use, the free tier is enough. Check exa.ai/pricing for current rates.

Can I filter by date or domain?

Yes. The API supports filters for include/exclude domains, published date ranges, and content type. The MCP server forwards these as query parameters.

Should I use Exa instead of Brave Search?

Use both. Brave is faster and broader; Exa is slower and deeper. For 'tell me what's currently happening with X', use Brave. For 'find me high-quality writing about X', use Exa. Claude can route between them automatically if both are installed.