Web search, built for AI agents
Exa is a search API built specifically for AI agents and applications, rather than for humans clicking through blue links. It returns clean, structured, token-efficient web content - search results, extracted page contents, and multi-step agent research - designed to be consumed directly by LLMs.
Published by the vendor themselves — not community forks.
Official Exa MCP server for web search and web crawling (TypeScript).
Official skills for the Exa API.
Official Exa Python SDK.
We raised $85M to build the search engine for AIs — Exa AI Labs
| Plan | Price | Included | Notable |
|---|---|---|---|
| Search | $7 / 1,000 requests | Pay-as-you-go, credit-metered | Fast/neural search, Multiple search depth options |
| Deep Search (lite) | $12 / 1,000 requests | — | |
| Deep Search (standard) | $12 / 1,000 requests | — | |
| Deep Search (reasoning) | $15 / 1,000 requests | — | Reasoning-augmented deep search |
| Contents | $1 / 1,000 pages (per content type) | — | Token-efficient content extraction, Highlights |
| Answer | $5 / 1,000 requests | — | Direct question answering grounded in search |
| Monitors | $15 / 1,000 requests | — | Scheduled/recurring search monitoring |
| Agent API | $0.012 - $1.00 per request (fixed-effort tiers: Minimal/Low/Medium/High/XHigh), or up to $5/run metered | — | Hosted multi-step web research agent, Component-based metered billing option |
| Enterprise | Contact sales | — | Volume discounts, Higher rate limits, SLAs |
Checked 2026-07-30 against their pricing page. Vendors change pricing without notice.
Exa (formerly Metaphor Systems) is a San Francisco-based AI research lab, founded by Will Bryk and Jeff Wang, building what it calls a search engine 'for AI agents' rather than for human browsers.
Engineering teams building AI agents, coding assistants, or research/enrichment products that need real-time, LLM-ready web data via API - from solo developers prototyping on free credits up to enterprise buyers needing SOC 2, zero-data-retention, and SLA guarantees.
Teams seeking a point-and-click GTM execution product should note that the official material retrieved emphasizes search, contents, answer, agent, and Websets surfaces rather than CRM or outbound sequencing workflows. This pass did not retrieve a broader non-technical application beyond Websets. Usage-based, per-endpoint pricing also requires buyers to forecast workload by endpoint.
APIs that answer a query with retrieved web content, built to be called by an agent rather than read by a person.
Category: Search and research APIs. Membership is decided by 2 stated tests, not by category tags — see them and the full comparison.
Appears in comparison shortlists
Yes - Exa is API-first.
New accounts get $20 in free credits (about 2,800 searches per the vendor), and the MCP server can be used without an API key at a rate-limited free tier.
Usage-based, metered per product: $7/1k search requests, $1/1k content pages, $5/1k answer requests, $12 - 15/1k deep-search requests, plus separate Agent API and Enterprise pricing. There is no flat monthly subscription on the public pricing page.
Deep Search is a higher-latency, higher-cost tier (lite/standard/reasoning) intended for more comprehensive, citation-backed research versus the faster default search endpoint.
A product for turning a natural-language query into a curated, enriched list of companies or people - targeted at sourcing/list-building use cases rather than single-query search.
The Enterprise plan page references SOC 2 Type II, Zero Data Retention, and SSO as available; the public vulnerability-disclosure policy confirms a formal security program but doesn't itself restate these certifications.
A first API call or MCP connection can happen in minutes with just an API key (or no key at all for basic MCP use); production integrations with structured outputs or the Agent API take longer to tune.
This page tells you whether Exa fits. Cheetah builds and runs the system it goes into — enrichment, sequencing, CRM and the joins between them.
Book a call →The directory stays free. There is nothing gated behind this.