which search api have you already used?

if you asked cursor a question today, maybe you used exa. if claude searched the web for you, maybe you used brave. if a langchain agent looked something up, that might have been tavily. you have probably used atleast one of five retrieval apis this week.

coding agentsassistants & chatagent frameworksvertical agentsplatforms & searchcursordevin (cognition)windsurfclinereplit agentclaude (anthropic)notion aizapier chatbotsretell voice agentslangchain / langgraphllamaindexn8nagnoopenclawclayharveyopendoorcredalhubspotmonday.comopenrouterdatabricksawssnowflake agentskagistanford ai playgroundexasemantic search, own index8 / 26 productsbraveindependent web index7 / 26 productstavilysearch + extract for frameworks4 / 26 productsfirecrawlscrape / crawl → llm-ready5 / 26 productsparallelsearch + deep research stack4 / 26 products
hover — or tap — any name to trace the dependency
a line means the relationship is publicly documented. every table row links to its source.
view as table
productretrieval apidocumented as
cursorexa, braveexa is its default web-search backend; brave also names it a customer source ↗
devin (cognition)exaexa lists cognition among production users source ↗
windsurfbravenamed by brave as an api customer source ↗
clinebravenamed by brave as an api customer source ↗
replit agentfirecrawlkeeps replit agent current with api docs and web content source ↗
claude (anthropic)braveclaude's web search is grounded with the brave index source ↗
notion aiexa, parallelnamed as a customer by both exa and parallel source ↗
zapier chatbotsfirecrawlcustom knowledge for customer chatbots via firecrawl source ↗
retell voice agentsfirecrawlturns customer docs into knowledge bases for ai phone agents source ↗
langchain / langgraphtavilytavily is the path-of-least-resistance search tool in langchain source ↗
llamaindextavilyfirst-party tavily integration source ↗
n8ntavilyfirst-party tavily integration source ↗
agnotavilyfirst-party tavily integration source ↗
openclawbrave~700k openclaw users signed up for the brave search api source ↗
clayparallelsales intelligence enrichment on parallel source ↗
harveyparallellegal research on parallel source ↗
opendoorparallelreal-estate operations on parallel source ↗
credalfirecrawlweb scraping at scale for enterprise ai agents source ↗
hubspotexanamed by exa as a production customer source ↗
monday.comexanamed by exa as a production customer source ↗
openrouterexaweb-search grounding for routed models source ↗
databricksexanamed by exa as a customer source ↗
awsexapublicly uses exa as a web-search backend source ↗
snowflake agentsbravebrave search api integrated for enterprise agentic search source ↗
kagibraverelies on brave as a core index provider source ↗
stanford ai playgroundfirecrawl~800 sources/day across 10,000+ domains via search + scrape source ↗

why you should care which one sits under your agent

your answers inherit its blind spots. an agent cannot cite what its retrieval layer never found. the model gets the credit, but the index sets the ceiling — same prompt, different retrieval stack, different report. that gap is exactly what the blind battles here measure.

it's a concentration risk. llms are frozen at their training cutoff, so every web-aware agent needs a retrieval layer — and google doesn't license its index for this. microsoft retired the bing search api in august 2025; the endpoint now returns http 410 and the azure replacement costs 40–483% more. that pushed nearly the entire agent ecosystem onto the five vendors above. a repricing or outage at any one of them degrades a large slice of ai products at once.

they're not interchangeable. brave is a raw independent index. exa is semantic search over its own crawl. tavily is the plumbing inside agent frameworks. firecrawl turns arbitrary pages into llm-ready markdown. parallel is a full agentic research stack — and the only one on this list that also ships a finished deep research product, which is why four parallel variants sit on our leaderboard while the others power the products around it.

openai, gemini, and perplexity built their own crawl-and-index stacks — they're vertically integrated. almost everyone else's "web-aware agent" is a composition of the five vendors above. when you compare deep research providers, you're partly comparing retrieval layers; when you pick a search api for your own agent, you're picking your ceiling.

methodology & sources

compiled july 2026 from vendor customer pages, engineering blogs, and press coverage. a line appears only when the vendor or the customer has publicly named the relationship — anonymous "teams at" claims are excluded. many products use more than one api; absence of a line means undocumented, not unused. every row in the map's table view links to the page documenting that specific relationship.