Tavily vs Exa: Which Search API Should Your Agent Use?
November 19, 2026

Tavily vs Exa comes down to one question: does your agent need to read a site, or find the right pages across the web? Both are web search APIs built for AI agents, both return LLM-ready content, and both ship an MCP server. They're built around different jobs.
The short version, as of November 2026:
- Pick Tavily if your agent searches, then extracts pages, crawls a whole site or maps its structure, and you want all of that behind one API key and one credit balance.
- Pick Exa if your agent sends natural-language queries and needs page text, highlights or summaries in the same call, or if you also want datasets, monitors and research runs from your search vendor.
- Test both on your own queries. Each vendor publishes its own speed and quality numbers. We don't repeat them as facts here, because they depend on your queries and they go stale quickly.
One ownership fact matters before you choose. Tavily is now part of Nebius. Exa's own site announces no ownership change. Everything below comes from each vendor's own docs and pricing pages, checked on November 19, 2026.
Quick Comparison
| Tavily | Exa | |
|---|---|---|
| What it is | "The web access layer for AI agents" | Web search API with its own crawler and index |
| Owner | Nebius (acquisition announced February 2026) | Exa (no ownership change announced on its site) |
| Core endpoints | Search, Extract, Crawl, Map, Research | Search, Contents, Answer, Monitors, Agent, Websets |
| Speed settings | Four search depths: ultra-fast, fast, basic, advanced | Four search types: lite, auto, deep-lite, deep |
| Results per search | Up to 20 | Up to 10 by default, up to 100 |
| What comes back | Ranked snippets per source, optional answer and full page | Ranked results with text, highlights or summaries |
| Site crawling | Yes, Crawl and Map endpoints | Subpage crawling through Contents |
| Datasets and monitoring | No listed endpoint | Websets and Monitors |
| Billing | Credits: monthly plans or pay as you go | Pay as you go per request, per product, no subscription |
| Agent tooling | Hosted MCP server, Python and JS SDKs, CLI | Hosted MCP server, Python and JS SDKs, agent skills |
Ownership: Tavily Joined Nebius, Exa Runs Its Own Index
On February 10, 2026, Tavily announced it was joining Nebius, the AI cloud company. In that post, Tavily's founder says "Nothing changes about how you use Tavily today" and that the API stays the same. Tavily's site now carries "Tavily by Nebius" branding.
For most teams, this changes nothing day to day. It matters if your company has rules about depending on a vendor owned by a cloud provider you don't use, or if you're choosing a long-term search layer and want to weigh where each roadmap is headed. The full story, and the other options it sent people looking at, is in our Tavily alternatives guide.
Exa runs its own crawler, ExaSearchBot, and its own index. Its docs say that, as of October 2026, the index "tracks over 2 trillion URLs and serves 100 billion pages." That's Exa's own figure, not an independent count.
Search: Depth Settings vs Search Types
Both APIs let you trade speed for quality per request. They slice it differently.
Tavily: four depths and a lot of filters
Tavily's Search endpoint reference lists four search_depth values:
| Depth | What the docs say | Credits |
|---|---|---|
ultra-fast | Lowest latency, one summary per URL | 1 |
fast | Lower latency, good relevance | 1 |
basic (default) | Balanced relevance and latency | 1 |
advanced | Highest relevance, higher latency | 2 |
Results come back as ranked sources, each with up to three snippets. The docs put those at about 500 characters for basic and fast, and about 800 for advanced. You can return up to 20 results.
Where Tavily stands out is control over which sources you get:
- Topics:
general,newsandfinance. News mode adds published dates automatically. - Domain lists: include up to 300 domains and exclude up to 150. A
prefermode boosts your listed domains without blocking everything else. - Dates and freshness: a
time_rangefilter, start and end dates, and a cache-age setting that can force a live fetch. - Other knobs: country boosting, language filters, exact phrase matching and an
auto_parametersmode that picks settings for you. The docs warn that auto mode can raise a request toadvanced, which costs 2 credits. - Optional extras: an LLM-generated answer (
include_answer) and the full page as markdown or text (include_raw_content).
Exa: search by description, with content in one call
Exa's search docs describe an API that "takes a natural language query and returns ranked web results with clean page content." Exa's homepage says it trains its own embedding models for retrieval, so you can write the query as a plain description of what you're looking for.
Exa's four search types, with the latencies its docs give:
| Type | Typical latency | Use it when |
|---|---|---|
lite | about 200 ms | Cost or speed matters most |
auto (default) | about 1 second | General use |
deep-lite | about 4 seconds | Light research and synthesis |
deep | 4 to 15 seconds | Multi-step search with stronger synthesis |
One detail worth knowing: lite never fetches a page live. It serves content from Exa's cache, and results with no cached copy are left out when you ask for text or highlights. Exa also offers data categories for source-specific search, such as companies, people, news and code.
Each result can carry full text, highlights or an AI summary. Searches return 10 results by default and up to 100 with numResults, with no pagination.
The practical difference: Tavily gives you more filters on which sources come back. Exa gives you more ways to shape what each result contains.
Beyond Search: Crawling vs Research Products
This is where the two differ most.
Tavily: built to read whole sites
Tavily's docs list five endpoints:
- Search for current sources
- Extract to pull clean content from one or more URLs
- Crawl to read many pages across a site
- Map to discover a site's structure
- Research for a cited report from multi-source research
Crawl and Map are Tavily's distinctive pair. If your agent needs to ingest a documentation site, follow links inside one domain, or list every page under a path, Tavily has a dedicated endpoint for each. Exa's Contents endpoint can also crawl subpages, up to 100 per request, and a subpageTarget term narrows which ones it picks. That covers "read the pricing and docs pages of this site", but it isn't a site map.
Exa: datasets, monitors and an agent
Exa's API has grown in a different direction. Beside Search and Contents, its pricing docs list:
- Answer for an LLM answer informed by Exa search results
- Monitors, which run an Exa search on a schedule and send new results to your webhook
- Agent for long-running research, list building and multi-hop enrichment, with fixed effort levels or metered runs
- Websets for building verified, enriched datasets from the web
If your product is "find every company that matches these criteria and keep the list fresh", Exa covers it from one account. Tavily's docs list no list-building or monitoring endpoint. Its Research endpoint writes cited reports.
| Job | Tavily | Exa |
|---|---|---|
| Read many pages on one site | Crawl | Contents with subpages |
| List a site's URLs | Map | Not a listed endpoint |
| Content for URLs you already have | Extract | Contents |
| Generated answer to a question | Search with include_answer, or Research | Answer |
| Multi-step research report | Research | Agent, or deep search |
| Build and refresh a list of entities | No | Websets, Monitors |
Pricing Model: Credits vs Per-Request Rates
Both bills scale with usage, but the units differ. Check both pricing pages before you budget, since rates change.
Tavily bills in credits. Its API credits page sets the exchange rate for each endpoint:
- Search: 1 credit, or 2 for
advanced - Extract: 1 credit per 5 successful URLs, or 2 for advanced extraction
- Map: 1 credit per 10 pages, or 2 with instructions
- Crawl: the Map cost plus the Extract cost
- Research: a range per request that depends on the model, from single digits up to a few hundred credits
Failed extractions and map requests aren't charged. You buy credits through monthly plans, where the per-credit price drops as the plan grows, or pay as you go. A monthly allowance of credits comes at no charge with no card. The plans are on Tavily's pricing page.
Exa bills per request, per product. Its pricing docs say there's no subscription and no minimum spend. Each search type has its own rate per thousand requests, with deep types costing more than lite and auto. Page contents are billed per page and per content type. Extra results beyond ten, and AI summaries, cost extra. Agent runs and Websets have their own meters. New accounts get a monthly credit balance that resets each month.
What this means in practice:
- A Tavily run is easy to count if you mostly search. It gets harder once one agent run mixes search, extract, crawl and research.
- An Exa run has several meters. A search that returns text and summaries for 20 results bills the search, the extra results and the summaries.
Either way, price a real agent run, not a single call. One user question can trigger five searches and ten page reads.
Security, Data Handling and Integrations
Both target production agents, so both cover the basics.
Tavily shows SOC 2, ISO 27001, GDPR and zero data retention badges on its homepage, links a Trust Center, and offers a Tavily EU option. Its homepage also says requests pass through layers meant to block PII leakage, prompt injection and malicious sources. Tavily's Nebius post says data policies stay unchanged and zero data retention remains central.
Exa lists Zero Data Retention as part of its annual Enterprise plan in its pricing docs, with custom pricing and invoicing through sales.
For wiring them into an agent:
| Tavily | Exa | |
|---|---|---|
| Hosted MCP server | Yes | Yes |
| SDKs | Python, JavaScript | Python, JavaScript |
| Other | A CLI, plus keyless calls to Search and Extract | An installable agent skill |
If you're adding either one to Claude Code or another MCP client, our list of the best MCP servers shows where search fits next to other tools.
Tavily vs Exa: Which Should You Pick
Choose Tavily if:
- Your agent needs to crawl or map sites, not just search them.
- You want search, extraction, crawling and research behind one credit balance.
- You filter heavily by domain, topic, country or date.
- Your security review wants SOC 2 and ISO 27001, or an EU option, listed on the vendor's own site.
Choose Exa if:
- Your queries read like plain descriptions of what you're looking for.
- You want text, highlights or summaries returned with each result.
- You build enriched datasets from the web, or you want scheduled searches sent to a webhook.
- You'd rather pay per request with no subscription or plan to manage.
Either works if you're prototyping a research agent. Both offer a monthly allowance at no charge, so run twenty of your real questions through each and compare the sources you get back.
If neither fits, we compare the wider field in Exa alternatives and the Tavily guide linked above. If your assistant already has search built in, you may not need either one. Our guide to Claude web search covers what the built-in tool handles.
Where Taku Fits
Tavily and Exa are building blocks. Someone still has to wire one into an agent and ship it. If you'd rather run a research tool someone has already built, that's where Taku comes in. Taku Apps and Sites, built by the best AI minds. Find one, run it, and keep it in your Mac Dock. Taku is in Beta, and the Mac app is available now.
Key Points
- Tavily is built around reading the web for an agent: search, extract, crawl, map and research, all billed in credits.
- Exa is built around natural-language search with content in the same call, plus answers, monitors, Websets and an agent.
- Tavily is now part of Nebius, and says its API hasn't changed. Exa runs its own crawler and index.
- Tavily has more filters per query. Exa has more control over what each result returns.
- Billing models differ: Tavily credits on monthly plans or pay as you go, Exa per request per product with no subscription.
- Speed and quality claims come from the vendors. Test both on your own queries.
FAQ
Is Tavily or Exa better for AI agents?
Neither wins across the board. Tavily is the better fit if your agent crawls sites or mixes search with extraction. Exa is the better fit if your agent sends natural-language queries and needs page content, datasets or monitoring. Run your own queries through both before you decide.
What is the main difference between Tavily and Exa?
Scope. Tavily covers search plus site crawling, mapping and extraction. Exa covers search plus page contents, answers, Websets and scheduled monitoring, on top of its own web index.
Who owns Tavily now?
Nebius. The acquisition was announced on February 10, 2026, and Tavily says the API and its data policies stay the same.
Does Exa have its own search index?
Yes. Exa runs its own crawler, ExaSearchBot, and says its index tracks over 2 trillion URLs as of October 2026. That figure comes from Exa.
Which is cheaper, Tavily or Exa?
It depends on your mix of calls. Tavily charges credits that vary by endpoint and depth. Exa charges per request for each product, plus extras for page contents, results beyond ten and summaries. Price one full agent run on each vendor's pricing page.
Can I use Tavily or Exa with Claude or Cursor?
Yes. Both run a hosted MCP server, so any MCP client can call them. Tavily's docs include setup commands for Claude Code, Cursor and Codex.