Best MCP servers for web research: Exa, Tavily, Brave, Firecrawl and more compared
Exa is the best MCP server for web research for most agents: one call returns search results with the page text, it starts without an account, and it was named first in 7 of 14 AI assistant answers we logged. For reading and crawling sites you already know, pick Firecrawl. We compared 7 servers on search versus extraction, free allowance, how calls are billed, setup, and how often AI assistants recommend them. Facts checked on 2026-10-01.
| Name | Best for | Made by | License | Notes | Checked |
|---|---|---|---|---|---|
| Exa | Search that returns page text in one call, keyless to start | Exa | MIT | Free plan | |
| Firecrawl | Scraping, mapping and crawling sites you already know | Firecrawl | MIT | Free plan | |
| Tavily | Search plus extract, map and crawl on one free allowance | Tavily | MIT | Free plan | |
| Brave Search | An independent index, news and local search, run locally | Brave | MIT | Free plan | |
| Perplexity | Cited answers and long research reports, paid per call | Perplexity | MIT | Paid only | |
| Jina AI | Reading pages and PDFs as Markdown, arXiv and SSRN search | Jina AI (part of Elastic) | Apache-2.0 | Free plan | |
| Apify | Sites that block plain fetching, such as Google Maps | Apify | MIT | Free plan |
What AI assistants recommend for web research
We asked "Which MCP servers should I connect to my AI agent for web research?" on 2026-10-01, in 14 runs across six engines, and counted every server each answer named and which it named first. Each cell reads named (named first).
| Server | ChatGPT, 3 runs | Gemini, 3 | Perplexity, 1 | Google AI Mode, 3 | Google AI Overviews, 1 | Grok, 3 | Named, of 14 | Named first |
|---|---|---|---|---|---|---|---|---|
| Exa | 3 (2) | 3 (2) | 1 (0) | 3 (1) | 1 (0) | 3 (2) | 14 | 7 |
| Firecrawl | 3 (1) | 3 (1) | 1 (1) | 3 (0) | 1 (0) | 3 (0) | 14 | 3 |
| Brave Search | 3 (0) | 3 (0) | 1 (0) | 3 (1) | 1 (1) | 3 (0) | 14 | 2 |
| Tavily | 2 (0) | 3 (0) | 1 (0) | 3 (1) | 0 | 2 (0) | 11 | 1 |
| Playwright | 2 (0) | 3 (0) | 1 (0) | 3 (0) | 1 (0) | 2 (0) | 12 | 0 |
| Fetch | 1 (0) | 3 (0) | 0 | 2 (0) | 0 | 3 (1) | 9 | 1 |
| Perplexity | 0 | 2 (0) | 1 (0) | 1 (0) | 0 | 2 (0) | 6 | 0 |
| Jina AI | 0 | 2 (0) | 0 | 0 | 0 | 0 | 2 | 0 |
| Parallel | 1 (0) | 0 | 1 (0) | 0 | 0 | 0 | 2 | 0 |
| Apify | 0 | 0 | 1 (0) | 0 | 0 | 0 | 1 | 0 |
The set of names was stable and the order was not: Google AI Mode put a different server first in each of its three runs (Exa, then Brave Search, then Tavily). Claude agreed on the set and differed on the order. In our Claude test, 15 runs on the same date, Brave Search, Exa, Tavily, Firecrawl, Fetch and Playwright were named every time, and the first pick split Tavily 6, Brave 4, Exa 3, Fetch 2. Two counting notes: one Grok answer was only partly logged, so its counts are a minimum; and one Gemini answer opened with a summary table led by Firecrawl, then a longer list led by Brave Search, and we counted Firecrawl as first. Full method in the field study.
Exa: best overall for agent web search
Exa is an official MCP server from Exa, a search engine built for machines, for agents that need to look something up and read the source in the same step. Its strength is that the default search tool returns clean page content with each hit, and a second tool reads any URL as Markdown, so the agent rarely needs a separate scraper. Exa hosts the server at mcp.exa.ai and it answers without an account at free rate limits; signed-in accounts get a monthly free credit and then pay per request (current prices). Its limit is heavy anonymous use: keyless calls hit rate limits quickly, and the agent_run research tool needs an account. Exa's docs cover ChatGPT, Codex, Claude, Grok Build, Gemini CLI and Cursor. The repo had 5,070 GitHub stars on 2026-10-01, and Exa was named in all 14 engine answers in our test. Details: Exa MCP server.
Firecrawl: best for scraping and crawling known sites
Firecrawl is the official MCP server for Firecrawl's web data API, for agents that must pull full pages, structured data or whole site sections rather than search snippets. Its strength is range: it scrapes a URL into Markdown or into JSON that matches your schema, maps every URL on a site, crawls within limits you set, parses PDFs, and can click and type on a page before reading it. A keyless hosted endpoint covers search, scrape and parse with daily limits. Its limit is cost at volume: crawl, map and agent need an API key, and every call spends credits, from a monthly allowance on the Free plan or a paid plan (current prices). Each interact call runs one turn, so a long browser session fits Playwright better. The MIT repo had 7,538 GitHub stars on 2026-10-01, and Perplexity named Firecrawl first, with Firecrawl's own list of best MCP servers among its top three search results. Details: Firecrawl MCP server.
Tavily: search and site tools on one free allowance
Tavily is the official MCP server for Tavily's search API, made for language models, for agents that want search, extraction and site crawling from one provider. Its strength is that one key covers five tools: search with depth and time-range options, extract for a list of URLs, map for a site's page list, crawl, and a research tool for longer questions. Defaults such as search depth can be set once for every call. The free Researcher plan gives a monthly credit allowance with no card, and past it you pay as you go (current prices). Its limit is that there is no keyless mode: every call needs an API key or OAuth sign-in, unlike Exa and Firecrawl. The MIT repo had 2,415 GitHub stars on 2026-10-01. Tavily was named in 11 of 14 engine answers in our test, and it was Claude's most frequent first pick, 6 of 15 runs. Details: Tavily MCP server.
Brave Search: an independent index, run on your machine
Brave Search is Brave's official MCP server for the Brave Search API, for people who want results from an index that is not Google's or Bing's, plus news, image, video and local search. Its strength is the spread of search types and a tool that returns pre-extracted text from the top results with token limits, so the agent gets readable sources in one call; Goggles let you re-rank results with your own rules. You can allow or block single tools with environment variables. Its limit is depth: the server returns results and extracted snippets, not whole pages or crawls, and it has no hosted endpoint, so it runs locally with npx or Docker. Every account gets a monthly free credit, and requests past it are billed (current prices). The MIT repo had 1,478 GitHub stars on 2026-10-01, and Brave Search was named in all 14 engine answers. Details: Brave Search MCP server.
Perplexity: cited answers and long research reports
Perplexity is the official MCP server for the Perplexity API Platform, for agents that should get an answer with citations, or a full research report, rather than a list of links. Its strength is four tools of rising depth: ranked search results, a quick cited answer, step-by-step reasoning, and a research tool that can run for minutes and stream its progress. It connects by OAuth or an API key, hosted or local, and Perplexity's README covers Claude Code, Codex, Cursor, VS Code and Windsurf. Its limit is cost: the pricing page lists no free tier, search is billed per request, and the answer tools are billed per token plus a fee per web search (current prices), so research runs can cost far more than a search. A Perplexity Pro chat plan does not cover the API. The MIT repo had 2,549 GitHub stars on 2026-10-01, and Perplexity was named in 6 of 14 engine answers. Details: Perplexity MCP server.
Jina AI: reading pages and searching papers
Jina AI's official MCP server connects agents to Jina's Reader, Search, Embeddings and Reranker APIs, for research that leans on documents and papers. Jina AI is now part of Elastic. Its strength is reading: any web page or PDF comes back as Markdown, or only the passages that answer your question, and search tools cover the web, arXiv and SSRN, with a PDF tool that pulls out tables and equations. Reranking sorts a long list of hits by relevance before the agent reads them. Reading works without a key at a rate limit, and each new API key comes with free tokens (Reader page, 2026-10-01). Its limit is that search, reranking and PDF extraction need a key, and the server is hosted only, with no local install. The Apache-2.0 repo had 869 GitHub stars on 2026-10-01. Jina was named in 2 of 14 engine answers, both from Gemini. Details: Jina AI MCP server.
Apify: ready-made scrapers for hard sites
Apify's official MCP server gives agents the scrapers in Apify Store, which Apify calls Actors, for research that needs data from sites that resist simple fetching, such as Google Maps, Instagram or search result pages. Its strength is choice: the agent can search the store, read an Actor's input schema and price, run it, and read the results as rows; two tools come preloaded, a RAG web browser and a web fetch with JavaScript rendering. Its limit is predictable cost: each Actor sets its own price per run or per result, so check an Actor's details before letting an agent run it in a loop, and you answer for how third-party scrapers are used. The Free plan includes a monthly usage allowance, and paid plans add more (current prices). The MIT repo had 9,213 GitHub stars on 2026-10-01, the most of the seven. Only Perplexity named Apify, once, as a specialist pick. Details: Apify MCP server.
Servers the engines named that we do not list
Fetch, the reference server in modelcontextprotocol/servers, was named in 9 of 14 engine answers and in every Claude run. It fetches a URL and returns its content as Markdown, runs locally, and needs no account; it does not search, so it pairs with one of the search servers above. Grok put Fetch first in one run. Parallel's search MCP, a hosted server that its README describes as free with no API key, was named twice, by ChatGPT and Perplexity. Playwright came up in 12 of 14 answers as the tool for pages that need a real browser; for that job see Playwright MCP alternatives.
How we compared
We read each server's GitHub README, hosted endpoint docs and pricing page on 2026-10-01 and took stars, licenses and free tiers from our item pages, checked the same day. The criteria were: whether the server searches, reads pages or crawls sites; what works without an account; whether there is a free allowance and how calls are billed; hosted or local setup; and the clients the publisher documents. The recommendation counts come from our six-engine field study and our Claude test, both run on 2026-10-01, counting names in each answer body. We did not run the same research task through each server, and we did not measure search quality, speed or how well pages were extracted. We do not list prices because they change often; read the publisher's pricing page before you commit.
How to choose
- If your agent mostly needs to look things up and read the sources, start with Exa.
- If you already know the sites and need full pages, JSON or a crawl, add Firecrawl.
- If you want search, extract and crawl under one free monthly allowance, choose Tavily.
- If you want an independent index, news or local search, and a server that runs on your machine, choose Brave Search.
- If you want a cited answer or a long report rather than raw results, and can pay per call, choose Perplexity; for papers and PDFs, Jina AI; for sites that block plain fetching, Apify.
Questions people ask
Which MCP server do AI assistants recommend for web research?
Exa, Firecrawl and Brave Search were named in all 14 answers we logged from ChatGPT, Gemini, Perplexity, Google AI Mode, Google AI Overviews and Grok on 2026-10-01, and Tavily in 11. Exa was named first most often, in 7 of 14. On Claude, over 15 runs the same day, Tavily was the most frequent first pick, in 6.
Is there a free MCP server for web search?
Yes, within limits. Exa's hosted server answers without an account at rate limits, Firecrawl's keyless endpoint covers search, scrape and parse with daily limits, and Jina reads pages without a key at a rate limit. Tavily and Brave Search both have a free monthly allowance with an account (pricing pages read 2026-10-01). Perplexity's API lists no free tier.
Should I connect a search server and a scraping server together?
Often, yes. A search server such as Exa, Tavily or Brave Search finds pages, and a scraper such as Firecrawl reads or crawls them in full. Grok's answers in our test suggested that pair, and Google AI Mode advised no more than three research servers at once to keep the agent's tool list short.