Exa

What is Exa?

Exa is a web search API built for AI agents. Its own homepage puts it plainly: "Web search, built for AI agents — one API for search, crawling, and research." Where a traditional search engine returns a page of blue links meant for a human to click, Exa returns clean, structured data meant for a language model to reason over.

The technical difference is neural search. Instead of keyword matching, Exa uses embeddings-based retrieval, which is why it's strong at the queries agents actually run — company research, people discovery, and finding code repositories. It's fast enough for the agent loop (results in under 180ms) and it's accurate: Exa scores 54.4% on the FRAMES benchmark, well ahead of Perplexity (44.5%) and Brave (21.6%). Teams behind serious agentic products — Cognition's Devin, Cursor, HubSpot, Monday.com, AWS, Databricks, and Groq — build on it.

Key Features

How to Use Exa

There are two easy on-ramps depending on how you build.

1. Grab an API key. Sign up, get a key from the dashboard, and you're on the free tier — 20,000 requests a month, no card.
2. Call the Search API. Send a query, choose how many results and whether you want highlights or full contents, and get back structured JSON with citations.
3. Or drop in the MCP server. If you build on Claude or another MCP agent, add Exa's MCP server and the agent can call web search as a native tool — no wrapper code.
4. Scale up when ready. Move to the Agent or Deep endpoints for autonomous, multi-step research, and add Monitors to watch the web over time.

The docs are genuinely good — clear endpoints, examples, and SDKs — so getting from key to first result takes minutes. The MCP path is the fastest if you're already in the agent tooling world.

Pricing Plans

Exa's pricing is usage-based and starts genuinely free — 20,000 requests a month before you pay a cent. After that, you pay per endpoint:

Pros and Cons

Best Alternatives

Exa competes in the fast-growing "search API for agents" space. Here's how it stacks up.

The short version: Tavily is simpler for basic RAG, and Firecrawl wins if you mainly need to crawl and scrape. But for accuracy, neural relevance, and research depth — plus a benchmark lead and enterprise adoption — Exa is the one to beat. It pairs especially well with the agent stacks we cover in our AI tools directory.

Frequently Asked Questions

❓ What is Exa?

A web search API built for AI agents. It uses neural, embeddings-based retrieval to return structured results with citations and highlights, so agents can pull current web data through one API.

❓ How much does Exa cost?

Free up to 20,000 requests/month, then metered: Search $7/1k, Contents $1/1k pages, Deep $12-15/1k, and Agent from $0.012 to $1.00 per request by effort. Enterprise is custom.

❓ How is Exa different from Google or Perplexity?

It's built for machines: neural search returns clean, cited excerpts instead of blue links, in under 180ms, and scores 54.4% on FRAMES versus Perplexity's 44.5%.

❓ Does Exa have an MCP server?

Yes — Exa ships an official MCP server, so Claude and other MCP agents can call Exa web search directly as a tool with no custom integration.

Final Verdict

4.7/5. Exa is the search layer the agent era actually needed. Neural retrieval, cited highlights that slash token costs, sub-180ms latency, a benchmark lead over Perplexity, and an official MCP server add up to the most agent-native search API on the market — and the roster of companies building on it (Cursor, Devin, HubSpot, AWS) is hard to argue with. It's developer-only and the metered pricing can climb on heavy Agent workloads, but neither is a real knock for the audience it's for. If you're building anything that needs to search the web intelligently, start with Exa's free tier — you'll likely stay.