Exa
Also known as: Metaphor Systems, Exa AI
Neural web search API built for AI agents and RAG that retrieves pages by meaning and returns token efficient, LLM ready contents.
Exa is a web search API built for AI rather than for people. Instead of matching keywords the way a traditional search engine does, it runs a neural index that retrieves pages by meaning, which suits the long, descriptive queries that agents and retrieval pipelines tend to produce. There is no consumer search box; Exa is called from code and returns clean, token efficient page contents and highlights that are ready to drop into a model's context. The company began as Metaphor Systems and rebranded to Exa in 2023.
Exa is really a family of retrieval endpoints. Standard search returns the best pages with their contents, while Instant trades a little depth for sub 200 millisecond latency for real time chat and voice agents, and Deep and Deep Reasoning modes handle research grade queries with structured outputs and grounded citations. A separate Contents endpoint pulls full page text as clean markdown, Answer returns cited answers, Monitors run searches on a schedule and push only new results to a webhook, and a newer Agent API runs multi step web research and list building. Exa is also strong on company, people, and code search.
For builders, Exa exposes structured outputs through an output schema with field level citations and confidence, plus filters for domain, path, subdomain, date, and user region. It ships native connectors for LangChain, LlamaIndex, and CrewAI, works as a Claude or OpenAI tool, and offers an official MCP Server, so it drops into most agent stacks without custom wrappers. On security, Exa supports customizable zero data retention, where queries and data can be purged automatically, along with access controls and single sign on for teams in regulated settings.
Pricing is pay per request with no seats and no monthly minimums. Every developer gets 1,000 free requests a month, then standard search is $7 per 1,000 requests, Instant is $5, Deep is $12 and Deep Reasoning $15, page contents and AI summaries are $1 per 1,000 pages, and Answer and Monitors are priced per 1,000 as well. A separate Websets product for B2B research starts at $49 a month, and enterprise plans add custom rate limits, dedicated indexing, and volume discounts. Exa raised an $85M Series B in early 2026 led by Lightspeed, with Nvidia and Y Combinator.
Vendor details
Canonical URL
https://exa.ai
Category
Agent infrastructure
Subcategory
Web search and retrieval
Funding status
Originally Metaphor Systems, rebranded to Exa in 2023. Y Combinator backed. Raised an $85M Series B in February 2026 led by Lightspeed, with Nvidia and Y Combinator participating. Used by teams including Notion and Vercel. Remains independent.
Company status
independent
Use cases & customers
Primary use cases
Target customers
Deployment options
Integrations
API first with native connectors for LangChain, LlamaIndex, and CrewAI, Claude and OpenAI tool use, and an official MCP Server. Structured outputs via outputSchema with field level citations, plus filters for domain, path, subdomain, date, and user region.
In practice
Your RAG pipeline returns keyword matched links that still need cleaning. You call Exa, which retrieves pages by meaning and returns clean, token efficient contents ready for the context window.
Your voice agent stalls because web search adds a second of latency. You switch to Exa Instant, which returns relevant results in under 200 milliseconds, so the conversation stays responsive.
You need to track competitor funding rounds without polling the web yourself. You set up an Exa Monitor that runs on a schedule and pushes only new, deduplicated results to your webhook.
Sources & related URLs
Related / legacy domains
Capability coverage
5.0 / 14 capabilities · 36%
| Integrations & Tool CallingNative connectors for LangChain, LlamaIndex, and CrewAI, Claude and OpenAI tool use, and an official MCP Server, serving as a retrieval tool agents call rather than a tool calling layer of its own. | Partial |
|---|---|
| Workflow OrchestrationThe Agent API orchestrates multi step web research and list building workflows, but Exa does not orchestrate general agent execution, sequencing, or branching. | Partial |
| Knowledge Grounding & RAGCore product. A neural retrieval engine purpose built for RAG and agent grounding, returning semantically relevant pages, token efficient contents, and web grounded answers with field level citations. | Full |
| Human Oversight & GuardrailsA retrieval API with no human approval, review, or guardrail features; grounded citations aid verification but are not an oversight layer. | Unable to verify |
| Security, Identity & GovernanceCustomizable zero data retention with automatic purging, access controls, and single sign on, but a lighter identity and governance surface than full enterprise platforms. | Partial |
| Observability & AuditabilityA data source rather than an observability tool; it provides usage visibility but no tracing or auditability of agent behavior. | Unable to verify |
| Memory & State PersistenceMonitors keep deduplication state across scheduled runs and the Agent API can build on an existing dataset, but Exa provides no agent memory or state persistence layer. | Unable to verify |
| Deployment & Data ResidencyA hosted SaaS API with no self hosted or on premises option, but customizable zero data retention lets regulated buyers control how long queries and data persist. | Partial |
| Prebuilt Agents, Templates & PacksOffers a packaged research agent via the Agent API, the Websets B2B research product, and specialized search modes for code, company, and people, but not general prebuilt agents or templates. | Partial |
| Triggers & Channel CoverageMonitors run searches on a schedule and deliver new, deduplicated results to a webhook, a real trigger and delivery channel, but Exa has no conversational channels or general agent invocation. | Partial |
| Model Flexibility & RoutingWorks as a tool for any external model and runs its own internal retrieval and reasoning models, but it offers no model selection or routing for the user's models. | Unable to verify |
| APIs, SDKs & MCP ExtensibilityAPI first design with Python and JavaScript SDKs, an official MCP Server, structured outputs via outputSchema, and framework connectors, built to be called from code. | Full |
| Testing, Debugging & OptimizationA retrieval API with no evaluation, testing, or optimization framework; grounded citations and confidence support verification but are not an eval suite. | Unable to verify |
| Browser & Computer UseNot applicable. Exa retrieves and crawls web content through an API and does not provide browser automation or computer use. | Unable to verify |
Pricing
Usage based · 1,000 free requests/mo · from $5/1k
Per 1,000 requests by search type, plus per 1,000 pages for contents and summaries
Included quota
1,000 free requests per month with full API access. Standard search includes contents for the first 10 results.
What is public
Exa publishes a full per request rate card: 1,000 free requests a month, then pay per request by search type, plus per page pricing for contents and summaries. The Websets product has fixed monthly tiers.
Billing mechanics
Billing is metered per 1,000 requests, priced by search type (Instant, Search, Deep, Deep Reasoning), with Contents and AI summaries billed per 1,000 pages and the Agent API billed per request by effort mode. There are no seats or monthly minimums. Websets is a separate fixed monthly subscription.
Cost watchouts
Additional results beyond the first 10, extra content types, AI summaries, and Agent enrichment add ons each carry their own per unit charges on top of the base search.
Variable cost rationale
Pure pay per request with no minimums keeps small workloads cheap, but cost scales linearly with query volume and search type, and heavy agent or research workloads using Deep, Reasoning, or Agent modes add up quickly.
Additional watchouts
Search and contents can be separate calls and separate charges, so a search plus contents workflow costs more than the headline search rate. Deep, Reasoning, and Agent modes are materially more expensive than standard search.
Overage / add-ons
Pay per request beyond the free tier: Instant $5, Search $7, Deep $12, Deep Reasoning $15 per 1,000 requests; Contents and AI summaries $1 per 1,000 pages; Answer $5 and Monitors $15 per 1,000. No seats or monthly minimums.
Sales call required
No — self-serve available
Free / trial
1,000 free requests per month, full API access, no credit card
Lowest paid plan
Pay per request; Instant from $5/1k, Search $7/1k
Commercial notes
Sold to developers and AI teams as a drop in retrieval layer, with bottom up adoption through the free tier and official framework and MCP integrations. Enterprise adds dedicated indexing, custom rate limits, security, and volume discounts.
Key ambiguities
Total cost depends on query volume, which search types are used, how many results and content types are pulled, and Agent effort levels.
Cancellation / refund
Self serve pay as you go with no minimums or seats; usage stops when you stop calling the API. Enterprise terms are contractual.
Support SLA / resale
Standard support on self serve; priority support, dedicated indexing, and invoicing on enterprise.
Missing data
Enterprise rates and volume discounts are custom. Exact Agent API costs depend on effort, compute, tool use, and enrichment add ons.
Related vendors
- AgentOps — Agent observability and reliability platform with broad model and…
- Agno — High-performance agent runtime and framework (formerly Phidata) with…
- Apify — Cloud platform for web scraping and automation with 45,000+ prebuilt…
- Arcade — Authenticated tool calling platform and MCP runtime that handles…
- Arize AI — AI observability and evaluation platform that traces, evaluates, and…
- Braintrust — AI evaluation and observability platform with self-serve pricing,…