Exa

Exa is a web search API built from the ground up for AI agents and LLM-powered applications. Founded as Metaphor Systems through Y Combinator (W21) and rebranded to Exa in January 2024, the San Francisco company operates a proprietary neural search index that ranks results by meaning rather than keyword frequency. CEO Will Bryk describes the core problem plainly: "Searching the internet should feel like navigating a grand library of knowledge. Today, searching the internet feels more like navigating a landfill." Exa raised an $85M Series B at a $700M valuation in September 2025, led by Benchmark with participation from Lightspeed, Y Combinator, and NVIDIA's NVentures arm.

The API offers several distinct endpoints: Search (semantic web retrieval with highlights extraction), Contents (clean full-page text from any URL), Answer (direct answers with citations), Find Similar (discover semantically related pages from a seed URL), and Monitors (real-time alerts on new matching content). On top of the API sits Websets, a research automation product launched February 26, 2025, that accepts plain-English research goals and deploys agent pipelines to verify thousands of sources and return structured tables of results. Native integrations ship for LangChain, LlamaIndex, CrewAI, and Claude's Model Context Protocol server, which had over 46,000 downloads by April 2025.

What Exa actually does in April 2026

Exa's core mechanism is an embeddings-first ranking system. Rather than matching query keywords against a page's keyword frequency (how Google's original PageRank worked), Exa encodes both the query and candidate pages into a shared vector space, then retrieves by semantic proximity. This means a query like "startup founders who left academia to commercialize protein folding" finds relevant pages even when those exact words never appear together on any single page.

The Search API returns results as structured JSON with URL, title, published date, author, and optional highlights: the specific sentences from the source document that match the query. This highlights feature, controlled via the highlights parameter, reduces the token load passed to downstream LLMs by 50-75% compared to feeding full page content, while improving RAG accuracy by roughly 10% according to Exa's internal benchmarks. Developers building retrieval-augmented generation systems get clean input without a secondary scraping call.

Beyond standard search, the /findsimilar endpoint takes a URL as input and returns semantically related pages, no query string needed. This powers discovery workflows: find competitors similar to a known startup, find articles similar to a reference piece, find job postings similar to a target role. The Monitors endpoint checks for new content matching a semantic query on a schedule, replacing manual Google Alerts with embedding-quality triggers.

The Websets product, accessed at websets.exa.ai, targets a different workflow: complex multi-source research that would otherwise require hours of manual tabulation. You define a goal in plain English ("CEOs of top 50 AI startups in the US by market cap"), and Exa deploys an agent pipeline that processes over 1,000 web pages, recursively verifies sources against your criteria, and outputs a structured table. The system found 20 times more correct results than Google and 10 times more than OpenAI Deep Research in Exa's own comparative testing on complex queries. Queries complete in minutes to tens of minutes, not seconds, so Websets is for research tasks rather than real-time retrieval.

Where Exa sits versus Tavily and Brave Search API

The developer search API market has three main contenders, each optimized differently.

Tavily was also purpose-built for AI agent consumption (not repurposed from a consumer product), which gives it ergonomic advantages: a search_depth parameter lets agents trade latency for result quality based on task urgency, and include_raw_content adds full-page extraction in a single call. But Tavily relies on keyword-based retrieval with LLM post-processing rather than neural ranking. Its P95 latency runs 3.8-4.5 seconds, versus Exa's 1.4-1.7 seconds. Tavily has no proprietary people, company, or code search sub-index. In independent AN Score agentic benchmarks, Exa scored 8.7 vs Tavily's 8.6, a narrow gap. Tavily was acquired by Nebius in February 2026 for $275M, which introduces some uncertainty about its independent roadmap. Exa's neural ranking wins on semantic retrieval tasks; Tavily's search_depth knob wins when an agent needs explicit quality-vs-speed tradeoffs.

Brave Search API takes a different angle entirely. Brave operates an independent web index of 30 billion-plus pages, entirely separate from Google's or Bing's infrastructure, and the API explicitly logs zero query data, making it the tool of choice for healthcare, legal, or government workloads where query confidentiality is a compliance requirement. Its free tier is more generous (2,000 queries per month vs Exa's 1,000). The tradeoff: Brave uses keyword-based ranking, so niche technical or conceptual queries return sparser results. In the same AN Score agentic benchmark, Brave scored 7.1, about 1.6 points below Exa. Choose Brave when the compliance posture or budget matters more than retrieval precision; choose Exa when semantic understanding is the primary requirement.

"Websets results are good because embedding models return actual startup homepages, while keyword-based engines return listicles." - HN user, Hacker News Launch thread, May 2025

What the agent integration workflow actually looks like

For a Python developer adding Exa to a LangChain agent, the integration is about a dozen lines: install the SDK (pip install exa-py), instantiate an Exa client with an API key, and use the built-in ExaSearchRetriever wrapper. For CrewAI, there is a dedicated @tool decorator pattern in the docs. The MCP server connects to Claude Desktop or any MCP-compatible client without code at all, surfacing web search, research, and code documentation search as native tools.

The pricing model is usage-based with no subscription required for the API tier. The free allocation of 1,000 requests per month covers prototyping and evaluation. At production scale, a team running 10,000 searches per month pays roughly $70 at the standard Search rate ($7 per 1,000). Startup and education projects can apply for $1,000 in free credits. Rate limits default to 10 QPS (600 requests per minute), which is adequate for most single-agent workflows but requires queuing logic for parallel multi-agent pipelines. Enterprise customers negotiate higher ceilings directly.

The Websets product is a separate credit-based subscription. The $49/month Starter tier provides 8,000 credits and 100 results per Webset. A single complex Webset with 100 verified results consumes a significant fraction of those credits, which catches new users off-guard. One HN user reported spending 750 of their 1,000 free credits on a single exploratory search. The recommendation from experienced users: use the developer dashboard to track credit consumption patterns before committing to a plan tier.

"The vast amount of the value of our whole search system comes from the retrieval, not from the LLM. Our customers often bring their own LLM and prompt." - Will Bryk, CEO, Sacra interview

Who Exa is built for

Exa targets AI engineers and developer teams as its primary audience. If you are building a RAG pipeline, an AI agent that needs to look things up, or a research automation product, Exa is the API that was designed for exactly that workload. The company's benchmark claims (54.4% accuracy on FRAMES vs competitors' 44.5% and 21.6%) and its roster of customers (Cursor, AWS, Databricks, Groq) suggest the quality bar is real.

Websets extends the value to non-engineer researchers who need structured, verified research at scale: sales teams building prospecting lists, analysts building competitive intelligence trackers, recruiting teams finding candidate pools. The plain-English interface removes the need to write API calls, but the Websets product is still slower and more expensive than a Google search for casual one-off lookups.

Within the developer audience, Exa is particularly strong for teams already using LangChain, LlamaIndex, or CrewAI, since native integrations mean no custom glue code. Claude MCP users get the fastest onboarding path: one click to add the connector in Claude Desktop, no terminal required.

What Exa is not

Exa is not a replacement for Google or Perplexity for everyday web browsing. There is no chat interface, no follow-up question flow, and no consumer UX. If you want an AI to search the web and answer a question conversationally, Perplexity Sonar or Claude with web search does that better.

Exa is also not the right tool for privacy-critical workloads. If your queries contain sensitive patient, legal, or financial information and you need a contractual guarantee that those queries are never logged, Brave Search API's zero-query-logging posture is more appropriate. Exa does not publish equivalent guarantees.

Websets is not suitable for real-time retrieval. Queries that return 50-100 verified results take minutes to tens of minutes, occasionally longer on complex criteria. If your agent loop needs a web search result in under two seconds, use the base Search API, not Websets.

Finally, Exa's credit system punishes intermittent users. The $49/month Websets plan works well only if you run searches consistently enough to consume the credits before the month resets. Occasional research teams may find the per-query API pricing more economical than a monthly subscription they partially burn.