Last updated: 2026-09-26
Exa is a neural web search API for AI agents and RAG pipelines, and HokAI rates it 4 out of 5: strong retrieval quality, a free Starter plan and an official MCP server, held back by default training use of submitted queries on self-serve plans and a low default search rate limit.
About Exa
Exa is an AI-native web search API built by Exa Labs (San Francisco, founded 2021) for large language models, autonomous agents, and retrieval-augmented generation pipelines. Harvard classmates Will Bryk (CEO) and Jeffrey Wang started the company roughly a year before ChatGPT launched, betting that AI needed a search engine built for machine consumption rather than human browsing.
The core mechanism is embeddings-based retrieval: Exa encodes every indexed page as a dense vector using its own trained neural networks, then runs nearest-neighbour lookups when a query arrives, surfacing pages whose vectors sit closest to the query. That captures conceptual proximity keyword search misses. Every request returns clean extracted text rather than raw HTML, cutting the token overhead of the downstream LLM call.
The Exa Search API is one endpoint, POST https://api.exa.ai/search, authenticated with an API key sent in an x-api-key or Authorization: Bearer header. Its type field sets the trade-off: auto is the default and balances quality against speed, fast trims latency for interactive work, instant targets sub-200ms agent loops, and three deep types are covered in the next paragraph. One call returns up to 100 results with title, URL, published date and author, filtered by category (company, people, news, publication, financial report or personal site), by publish date, and by up to 1,200 included or excluded domains. Page text, highlights and summaries come back in the same request through the contents option, and every response reports its own dollar cost. Separate endpoints handle /contents for URLs you already hold, /answer for a cited answer to a question, Monitors for scheduled searches delivered by webhook, and the Agent API.
Exa sells deep research two ways. Deep Search stays on the /search endpoint: deep-lite returns synthesised results at a consistent 4-second latency, deep runs multi-step research with synthesis, and deep-reasoning adds stronger reasoning for analysis and decision work. That last type is also the replacement for the old /research endpoint, which Exa retired on 1 April 2026 according to its changelog. Exa Agent, released on 16 June 2026, is the asynchronous route: you create a run with POST /agent/runs, then stream it or poll it, and it returns a written answer, JSON validated against an outputSchema you supply, and the citations behind both. Five fixed effort modes run from minimal to xhigh. auto is metered with a default cap of $5 a run, and ultra, added on 24 September 2026 for large list building and exhaustive multi-source research, is metered with a default cap of $20 that the caller can raise or lower.
There are three account plans. Starter is free with no payment method and includes every endpoint, all four indexes (web, people, company and scholarly works, the last expanded by 350 million publications on 23 July 2026), the MCP server, a Claude Connector and a ChatGPT plugin. Developer is pay-as-you-go and lifts search throughput to 25 queries per second. Enterprise adds zero data retention, HIPAA, SSO and SCIM, SLAs, custom indexes and larger result sets per search. Websets, the no-code dataset product, is a separate subscription. Exact rates for all of it sit in the Pricing section below.
Exa plugs into an agent stack rather than replacing one. The official MCP server connects it to Claude and Cursor, which Exa names as a customer using it to fetch current documentation, and the Python and JavaScript SDKs sit comfortably under LangChain or LlamaIndex retrieval code. For the wider field, browse the AI search and retrieval infrastructure directory or read HokAI's guide to retrieval infrastructure tools. The neighbouring API choices are weighed in Algolia vs You.com and Context.dev vs You.com, and the 2026 AI tool ecosystem guide places Exa in the wider stack.
Who should not buy it: if a person, not a program, reads the result, Perplexity gives a cited answer with no integration work. If the job is searching your own catalogue or docs, Algolia is the right tool, because Exa only searches the public web. On the API side, You.com and Context.dev are the head-to-heads worth running against your own queries before committing. Also read the privacy terms before sending anything sensitive: on Starter and Developer, Exa may use submitted queries to improve and train its models, and zero data retention is sold only on Enterprise. Exa's Vanta trust centre carries SOC 2 Type II and a HIPAA attestation dated 17 June 2026, with the audit reports behind an access request.
Screenshots
Pricing
Exa runs two pricing systems. ai/pricing on 26 September 2026. Search is $7 per 1,000 requests for up to 10 results, plus $1 per 1,000 for each result beyond ten.
Answer is $5 per 1,000. Deep Search is $12 per 1,000 for deep-lite and deep, and $15 per 1,000 for deep-reasoning. Monitors are $15 per 1,000.
Contents is $1 per 1,000 pages per content type, and AI page summaries add $1 per 1,000 pages, so a search that also pulls full page text lands near $8 per 1,000. 005 per search tool call instead. 07, billed separately.
The free Starter plan adds $10 in credits every month plus a $10 onboarding bonus and asks for no payment method; Developer bills the same rates pay-as-you-go. Websets is a separate subscription at $49/month for 8,000 credits or $449/month for 100,000 credits, with a custom Enterprise tier. Enterprise API pricing is custom, with volume discounts and invoice billing.
| Tier | Monthly price | What it includes |
|---|---|---|
| Starter (Free) | Free | $10 monthly credit and a one-time $10 sign-up bonus, no payment method, every endpoint and all four indexes, 10 Search QPS, 50 agent concurrency |
| Websets Starter | $49/mo | 8,000 credits a month, up to 100 results per Webset |
| Websets Pro | $449/mo | 100,000 credits a month, up to 1,000 results per Webset |
| Developer (Pay as you go) | Pay as you go | Per-request billing by card or bank, standard email support, SOC 2 Type II, up to 25 Search QPS, self-serve Exa Connect providers |
| Enterprise | Custom | Custom MSA and DPA, zero data retention, SOC 2 Type II and HIPAA, SSO and SCIM, SLAs, custom QPS, up to 1,000 results per search |
Feature Comparison by Tier
| Feature | Starter (Free) | Websets Starter | Websets Pro | Developer (Pay as you go) | Enterprise |
|---|---|---|---|---|---|
| Price | $0, monthly $10 credit, $10 sign-up bonus | $49/month | $449/month | Usage only, no commitment | Custom |
| Payment method required | No | Yes | Yes (free trial offered) | Card or bank account | Monthly invoicing |
| Search rate limit | 10 Search QPS | — | — | Up to 25 Search QPS | Custom Search QPS |
| Agent concurrency | 50 | — | — | 50 | Custom |
| Credits / results | All endpoints, all four indexes | 8,000 credits/mo, 100 results per Webset, 2 seats | 100,000 credits/mo, 1,000 results per Webset, 10 seats | All endpoints, all four indexes | Up to 1,000 results per search, 5,000 per Webset |
| Zero data retention | — | — | — | — | ✓ |
| Compliance | — | — | — | SOC 2 Type II | SOC 2 Type II and HIPAA, custom MSA and DPA |
| SSO / SCIM | — | — | — | — | ✓ |
| Support | — | Standard | Standard | Standard email | Dedicated Slack channel and SLAs |
| MCP server, Claude Connector, ChatGPT plugin | ✓ | — | — | ✓ | ✓ |
| Exa Connect data providers | — | — | — | Self-serve | Premium providers, bring your own |
Key Features
- Search API with six search types: `POST /search` accepts auto, fast, instant, deep-lite, deep or deep-reasoning, and filters by category, publish date and up to 1,200 domains.
- Deep Search and Agent for deep research: Deep Search answers on the search endpoint (deep-lite at a consistent 4-second latency), while Agent runs asynchronously at /agent/runs across seven effort modes with schema-validated JSON output and citations.
- Four indexes, not one web crawl: Web, people, company and scholarly-works indexes are all open on the free Starter plan, and the publication category gained 350 million publications on 23 July 2026, so agents query a pre-filtered vertical instead of generic open-web results.
- Contents: clean text instead of HTML: Returns parsed page text, query-relevant highlights and AI summaries for URLs you already hold, with configurable livecrawl and a much higher default query rate than the search endpoint.
- MCP server, Claude Connector and ChatGPT plugin: A free Model Context Protocol server plus a Claude Connector, a ChatGPT plugin and 50+ integrations, all on the free tier, so Claude, Cursor and VS Code reach Exa search without custom API glue.
- Monitors and Exa Connect: Monitors run saved searches on a daily or weekly cadence and push new results by webhook, and Exa Connect, launched 24 June 2026, routes Agent requests to data providers such as Similarweb and Fiber.ai.
Pros
- The free Starter plan is usable for real evaluation and needs no payment method: monthly credits, every endpoint and all four indexes rather than a crippled preview
- Official integrations cost nothing extra: Claude, Cursor and VS Code reach Exa through its MCP server and ChatGPT through a plugin, with no custom REST wrapper and no paid plan needed
- Pricing is published per product and every API response reports its own dollar cost, so spend can be modelled before writing code and tracked per call afterwards
- Runs its own neural index rather than reselling a third-party search engine, and offers people, company and scholarly-works indexes alongside the open web for knowledge-intensive agents
Cons
- On Starter and Developer, Exa's privacy policy lets it use submitted queries to train and fine-tune its models. Zero data retention is an Enterprise plan feature, so the queries an agent sends are training input by default
- Content extraction is billed on top of search, per page and per content type, and summaries are billed again, so a search that returns full text plus a summary costs noticeably more than the headline search rate
- The default search rate limit is low and pay-as-you-go only raises it to 25 QPS; anything beyond that needs an Enterprise conversation, so a high-concurrency pipeline still has to build queuing and retry logic or talk to sales
- It is an API, not a product: there is no interface for a non-technical team. The one no-code surface, Websets, is a separate monthly subscription rather than part of the API plan
Product Information
- Cloud
- Yes
- Self-Hosted
- No
- On-Premise
- No
- Languages
- English, Spanish, French, German, Japanese, Chinese, Portuguese, Arabic
- Training
- Documentation (exa.ai/docs), Python SDK examples, JavaScript SDK examples, Community Discord
Data Handling
- Training-data policy
- Exa's privacy policy states that "Query Data is used to improve our products and technology, including by training and fine-tuning models" — so on the Starter and Developer plans, queries are training input by default. Zero data retention is offered, but Exa's pricing page lists it as an Enterprise plan feature alongside a custom MSA and DPA, not as a default or a self-serve toggle. The privacy policy also says Exa does not monitor queries for personal information and advises users not to submit personal data through query fields, directing them to designated contact methods instead. No fixed retention period is published; the trust centre lists "Data retention procedures established" and "Customer data deleted upon leaving" as continuously monitored controls, and names a subprocessor list and a data processing addendum among its published resources. California residents are pointed to a Delete Act disclosure for consumer data requests.
- Compliance
- SOC 2 Type II and HIPAA · both listed on Exa's public Vanta trust centre at trust.exa.ai. Published artefacts include an Exa Labs Inc. HIPAA Attestation dated 2026-06-17 · the SOC 2 Type 2 report · a 2026 SOC 2 bridge letter · a data processing addendum · a penetration testing report and a subprocessor list — the audit reports and pentest are behind an access request. HIPAA appears as an Enterprise plan line alongside SOC 2 Type II on Exa's pricing page. Monitored controls include data encryption · restricted encryption key access and penetration testing.
Frequently Asked Questions
How much do you pay for Exa?
Exa's Search API charges $7 for every 1,000 requests returning up to ten results, according to its pricing page on 26 September 2026. Answer calls are $5 per 1,000, Deep Search runs $12 per 1,000 (deep-lite and deep) or $15 per 1,000 (deep-reasoning), and pulling page text adds $1 per 1,000 pages. Agent runs at a fixed effort range from $0.012 to $1.00 each. Websets starts at $49 a month, and Enterprise is quoted by sales.
What do you get on Exa's free tier?
The free Starter plan gives a $10 credit each month and a one-off $10 onboarding bonus, with no payment method needed. At the $7-per-1,000 search rate the monthly credit covers roughly 1,400 basic searches. It includes every endpoint, all four indexes and the MCP server.
What should you use instead of Exa?
Tavily is the closest agent-first rival and leans on LangChain and LlamaIndex integrations; Nebius agreed to buy it on 10 February 2026 and it still runs under its own brand. You.com sells a search API for grounding language models. Perplexity's Sonar API returns a cited answer in a single call. Algolia fits when you are searching your own content rather than the open web.
What separates Exa from Tavily?
Exa's own comparison page reports 81% against Tavily's 71% on WebWalker multi-hop retrieval, but that is a vendor-run evaluation, so test both on your own queries. Exa runs its own neural index with people, company and publication verticals and ships an official MCP server for Claude and Cursor. Tavily integrates more tightly with LangChain out of the box and has been owned by Nebius since 2026.
How long does it take to get going with Exa?
Minutes. Sign up at exa.ai with no card, copy an API key, then send POST https://api.exa.ai/search with the key in an x-api-key header and a JSON body holding your query. The type defaults to auto; add contents with text set to true to get page text in the same call. The exa-py and exa-js SDKs wrap the same endpoint, and the free MCP server exposes it inside Claude or Cursor.
Top Alternatives
- Perplexity AI: Perplexity answers a question directly with citations in a conversational interface; Exa returns structured, programmatic search results built for an AI agent to consume. Pick Perplexity when a person is reading the answer, Exa when a program is.
- Algolia: Algolia indexes and searches your own content at millisecond speed and is the right answer for site or in-app search. Exa searches the open web for research and agent grounding, and cannot search a dataset you have not published.
- You.com: You.com sells both a consumer assistant and a search API for LLM grounding, so it is the closer head-to-head on the API side. Exa's differentiator is running its own neural index with dedicated people, company and scholarly-works verticals, plus per-run agent pricing; check both against your own retrieval evaluation rather than a published benchmark, since each vendor's numbers come from its own testing.
HokAI guides covering Exa
- Best AI Search & Retrieval Infrastructure Tools in 2026: AI search and retrieval infrastructure is four purchases, not one: web search, document parsing, vector storage, and agent memory. Compare eight tools by layer.
- Four Lanes, 95 Guides, and Not One Outside Its Band: HokAI sorts its guides into four lanes: comparison, buyer's guide, news analysis and Inside HokAI. All 95 sit inside the word band their own lane declares.
- How HokAI Verifies a Listing Before It Goes Live: We pulled all 656 HokAI listing records from the public API on 18 August 2026 and counted the receipts: 39,161 fields and dated check stamps you can audit.
- Algolia and You.com Both Call Themselves AI Search Infrastructure. They Aren't Rivals.: Algolia and You.com both market themselves as AI search infrastructure in 2026, but Algolia indexes data you own while You.com searches the live web for agents.
- Context.dev vs You.com: Which Should You Use in 2026?: Both now sell web search and page extraction in 2026. Context.dev bundles brand data for $25/month; You.com alone has a published search-accuracy benchmark.
- The AI Tool Ecosystem in 2026: Buy the Meter, Not the Category: Four meters set your AI infrastructure bill: tokens, machine seconds, vendor credits, seats. Modal, Replicate, E2B and Together AI priced August 2026.