Exa

Exa is an embeddings-first web search API engineered specifically for LLMs and AI agents — instead of matching keywords, it retrieves the web by semantic similarity. We rate it 78/100 — an excellent neural retrieval engine for AI builders, but consumption pricing and a smaller index than Google mean it's not the right default for every team.

What is Exa?

Exa is a web search API designed from the ground up for AI consumers, not human readers. Founded in 2021 by Will Bryk and Jeff Wang under the name Metaphor Systems, the company went through Y Combinator's Winter 2022 batch and rebranded to Exa in January 2024. Headquartered in San Francisco, it has raised over $111M to date at a reported $700M valuation.

The core insight: pretraining LLMs and indexing search engines are remarkably similar processes. So Exa built a search engine using its own embeddings model and Neural PageRank — letting agents query the web with natural-language descriptions like "a small open-source library that does X" and get back genuinely relevant URLs, even single-star GitHub repos that classic keyword search would never surface.

Key Features of Exa

• Neural search: An embeddings-based retrieval engine that understands intent — searching by meaning rather than tokens. Particularly strong on long-tail queries where keyword search returns garbage.
• Auto + keyword modes: Three switchable modes (neural, keyword, auto) so you can fall back to traditional matching when you need a known string.
• Contents API: Fetches cleaned page text, highlights, and AI-generated summaries alongside results. As of March 2026, contents for the first 10 results per request are included free.
• Exa Deep & Deep-Reasoning: Agentic search modes that iteratively refine queries and reason over multiple pages — priced higher but useful for research workflows.
• Websets: A no-code product for non-developers to assemble curated lists of pages from the web with structured filters.
• SDKs & integrations: Official Python and TypeScript SDKs, plus first-class support in LangChain, LlamaIndex, and via the Exa MCP server.

What Users Say About Exa

Developer sentiment is largely positive on the technical side. On Hacker News and the Latent Space podcast, builders consistently praise Exa for surfacing semantically relevant results that Google's API misses — one frequently cited example is finding obscure single-star GitHub repositories from a vague natural-language description. Cursor and AWS publicly use Exa as their default web-search backend, which is a strong signal of production-readiness.

The criticism is just as consistent. Reddit and Product Hunt commenters flag that consumption pricing scales steeply — agents running thousands of queries a day rack up costs that are hard to forecast. Reviewers also note Exa's index is meaningfully smaller than Google's, so for breadth-heavy use cases (news monitoring, brand listening) you can hit coverage gaps. Several developers said they pair Exa with a separate scraping tool because content extraction, while improved, still misses some pages cleanly.

Exa Pricing

Exa runs on usage-based pricing with a generous free tier and a single Pro subscription. Effective March 2026, contents for the first 10 results are bundled into search calls — making most workloads roughly 30% cheaper than the prior model.

Plan	Price	Key Limits
Free	$0/month	$10 in free credits, ~1,000 searches/month
Pro (API)	From $40/month	$7/1K searches, $12/1K Deep, $15/1K Deep-Reasoning, $1/1K extra page contents
Websets	$49/month	8,000 credits, no-code list-building UI
Enterprise	Contact	Custom volume, SLAs, dedicated support

Who Should Use Exa?

Best for: AI engineers building research agents, RAG pipelines, lead-generation crawlers, and "find me X on the web" features where semantic relevance beats raw keyword recall. Solid fit for startups and teams already using LangChain, LlamaIndex, or MCP-based agents.

Not ideal for: High-volume consumer search products where forecasting cost is critical, or workloads that need Google-scale index breadth (real-time news monitoring across long-tail publishers, e-commerce SKU search across the entire web).

Pros and Cons

Pros:

• Best-in-class neural retrieval — finds relevant pages keyword search can't reach.
• Production-grade — Cursor and AWS run on it, so the reliability story is real.
• Free tier with $10 credit and ~1K searches/month is enough to fully prototype.
• Strong agentic features (Deep, Deep-Reasoning, Websets) for non-trivial research.
• Clean SDKs and first-class MCP server make integration genuinely fast.

Cons:

• Consumption pricing scales unpredictably — easy to blow through credits with an unsupervised agent.
• Index is smaller than Google's; long-tail and very recent content can be missed.
• Content extraction sometimes still requires a separate scraper for messy sites.
• No bundled discount for higher commitment tiers — pricing is mostly linear.

Alternatives to Exa

The closest alternatives are Tavily (cheaper, optimized for agent search but with weaker semantic ranking), Linkup (newer, French startup focused on LLM-friendly search — see our Linkup review), and Perplexity's Sonar API (good for question-answering but more opinionated and less granular). For pure breadth, Google or Bing search APIs still win on coverage but lose on agent-friendliness.

Verdict: Is Exa Worth It?

Yes — if your use case is an AI agent or RAG system that needs semantically relevant results, Exa is the most technically differentiated search API on the market today. The free tier is enough to prove value before committing, and the $7/1K pricing is fair for the quality you get. We'd skip it only if your workload is dominated by simple lookups where a cheaper Tavily-style API will do, or if you absolutely need Google-scale index breadth. Final rating: 78/100.

Frequently Asked Questions

Is Exa free?

Exa offers a free tier with $10 in credits — roughly 1,000 searches per month. Paid usage starts at $7 per 1,000 searches, and Pro plans begin at $40/month.

What platforms does Exa support?

Exa is an HTTP API with official SDKs for Python and TypeScript, plus a CLI and an MCP server for use inside agent frameworks like LangChain, LlamaIndex, and Claude.

How does Exa compare to Tavily?

Exa has stronger semantic ranking and a more sophisticated neural index, while Tavily is generally cheaper and tuned for simple agent web search. Most teams pick Exa when retrieval quality matters most and Tavily when cost-per-query is the priority.

Is Exa open source?

The Exa search engine itself is closed source, but Exa publishes open-source SDKs at github.com/exa-labs, including the official Python client.

Who uses Exa in production?

Publicly disclosed customers include Cursor (the AI code editor) and AWS, alongside many AI startups building agentic research and lead-gen products.