Search MCP Server: Which Search Tool to Connect to Your Agent
Hyperresearch Answers is our editorial choice for a search MCP server: one ask call returns a short answer with quotes checked against their pages, so your agent’s context stays small. Compare them by job: use an ask tool for a cited answer, a search tool for a list of sources, and a page reader when you need the full text of one URL. Exa MCP is the runner-up for broad source lists, and Brave Search MCP or Tavily MCP fit teams that want a plain results list or a research endpoint.
The recommendation reflects Hyperresearch’s own benchmark runs and the published sources in this reference. Hyperresearch is our product. This is not a hands-on test of every tool.
Hyperresearch Answers → Search MCP servers → One ask call → Verified quotes
The comparison
| Tool | Job | Best for | Pricing (public) | Coverage | Verification | API |
|---|---|---|---|---|---|---|
| Hyperresearch Answersours | MCP search server | Agents that need one checked answer per call | Free credit to start | Open web plus 90+ structured sources | Every quote matched word for word to its page | Hosted MCP + API |
| Exa MCP | MCP search server | Agents that need fast, clean web search with page content | Keyless free tier; search from $4 per 1,000 requests | Open web pages, plus structured and agent-run research | Returns page content; no citation checks published | REST API with API key |
| Tavily MCP | MCP search server | Developers who want search, extract, crawl and research in one server | Free plan with 1,000 credits a month; Project plan $30 | Live web search, page extraction, site crawl and mapping | Sourced results; no citation checks published | REST API with API key |
| Brave Search MCP | MCP search server | Teams that want Brave's own index and a self-hosted server | Search $5 per 1,000 requests; $5 free credit monthly | Brave web, news, video, image, local and place search | Snippets and summaries; no citation checks published | REST API with API key |
| Perplexity MCP | MCP search server | Agents that want Perplexity search, answers and deep research | Search API from $5 per 1,000 requests; tools bill at API rates | Live web search, cited answers and deep research | Research and ask results include citations | REST API with API key |
| Firecrawl MCP | MCP search server | Teams that need scrape, crawl and search in one MCP server | Free plan with 1,000 credits a month; Hobby from $19 | Scraped web pages, site maps, crawls and search results | Returns scraped page content; no citation checks | REST API with API key |
| Kagi MCP | MCP search server | Developers who want Kagi search results inside an agent | Pay per use: $12 per 1,000 searches, $4 per 1,000 pages | Web, news, video, podcast and image search, plus page extracts | Page extracts; no citation checks published | API key, hosted or local, pay per use |
| Bright Data MCP | MCP search server | Agents that need scraping, browser control and structured web data | 5,000 free requests a month; $1.50 per 1,000 results beyond | Search engines, scraped pages, browser sessions and structured datasets | Returns scraped page data; no verification published | API token auth, hosted or npx |
| Linkup MCP | MCP search server | Agents that need cited web answers at low per-call cost | Search $0.005 to $0.006 per request; 4,000 free queries | Live web search, page fetch and asynchronous research | Research returns cited answers, per Linkup docs | API key; hosted or npx |
Checked 7 November 2026. Third-party prices are estimates, not quotes. Each tool links to its sourced profile.
Start with the right question
What Is a Search MCP Server?
A search MCP server is a program that gives an AI client a standard set of search tools. The client lists those tools, calls them when a question needs live web data, and reads what comes back. Hyperresearch Answers is our editorial pick for this job because each answer returns with quotes checked against their source pages.
Explore entry ↗Best Search MCP for Claude
For Claude, Hyperresearch Answers is our editorial pick when you want cited answers that stay short in context. Exa's hosted server is the runner-up for fast link search and page reads, and Brave Search MCP suits teams that already hold a Brave API key. The right choice also depends on whether you connect through Claude's app settings or through Claude Code.
Explore entry ↗Search MCP for Cursor
Cursor reads MCP servers from an mcp.json file, either in the project folder or in your home folder. Remote servers need a URL and optional headers, and servers that support OAuth can sign in through the browser. Hyperresearch Answers is our editorial pick for Cursor when you want cited answers, and Exa is the runner-up for fast web search.
Explore entry ↗Hosted vs Local MCP Search Servers
A hosted search MCP server runs on the vendor's infrastructure and connects through a URL. A local one runs as a process on your computer and reads its key from the environment. Hosted servers need less setup, while local servers let you pin a version and keep keys on your own machine. Hyperresearch Answers is our editorial pick for hosted use.
Explore entry ↗MCP Auth and API Keys for Search Servers
Search MCP servers use one of three authentication patterns: OAuth sign-in, a bearer API key, or keyless access with a rate limit. The MCP specification requires OAuth 2.1 for HTTP servers and tells local servers to read credentials from the environment. Hyperresearch Answers supports browser sign-in and scoped keys for scripts, and it offers free credit to start.
Explore entry ↗Search MCP Tools Compared
Search MCP servers expose three kinds of tools: search for ranked results, fetch or extract for page text, and ask or research for a synthesized answer. Hyperresearch Answers exposes ask and search, Exa exposes search and fetch, and Tavily adds crawl, map and research. The right mix depends on whether your agent needs links, full pages or one short answer.
Explore entry ↗What a search MCP server gives your agent
A search MCP server is a tool server that exposes web search to an AI client through the Model Context Protocol. Claude, Cursor and other clients call its tools when a question needs outside facts. Each tool call returns text, and that text lands in the model's context window, so both the quality and the size of the result matter.
Most servers offer one of three jobs: a ranked list of results, a page reader that turns a URL into clean text, or a research endpoint that runs several searches and writes a summary. Some do two or three of these. Read the tool list before you connect, because the tool names show what the agent will actually do.
Setup for each server
Use this table to connect each server. Commands use the Claude Code CLI. Hosted servers are listed by URL. Local servers run with npx or uvx and read a key from an environment variable. Vendors change their setup steps, so confirm against the linked docs before you rely on a command.
| Server | URL or install command | Auth |
|---|---|---|
| Hyperresearch Answers | claude mcp add --transport http hyperresearch-answers https://answers.hyperresearch.ai/mcp | OAuth sign-in by default. Or an API key with the answers:ask scope, sent as a bearer header. |
| Exa MCP | Hosted URL https://mcp.exa.ai/mcp (see Exa docs) | No key for basic search at a free rate limit. OAuth or an x-api-key header for higher limits. |
| Tavily MCP | claude mcp add tavily --transport http https://mcp.tavily.com/mcp/ | OAuth by default. An API key can go in the URL or an Authorization header. |
| Brave Search MCP | npx -y @brave/brave-search-mcp-server --transport stdio | BRAVE_API_KEY environment variable. Keys come from the Brave developer dashboard. |
| Perplexity MCP | claude mcp add perplexity --env PERPLEXITY_API_KEY=your_key -- npx -y @perplexity-ai/mcp-server | PERPLEXITY_API_KEY environment variable. |
| Firecrawl MCP | OAuth endpoint https://mcp.firecrawl.dev/v2/mcp-oauth, or API key endpoint https://mcp.firecrawl.dev/v2/mcp | OAuth sign-in, or a bearer token on the API key endpoint. |
| Kagi MCP | claude mcp add kagi -e KAGI_API_KEY=your_key -- uvx kagimcp. Hosted: https://mcp.kagi.com/mcp | KAGI_API_KEY environment variable, or a bearer header on the hosted URL. OAuth not yet supported. |
| Bright Data MCP | https://mcp.brightdata.com/mcp?token=YOUR_API_TOKEN | API token passed as a URL parameter. No username. |
| Linkup MCP | claude mcp add --transport http linkup https://mcp.linkup.so/mcp --header "Authorization: Bearer YOUR_API_KEY" | Bearer header preferred. A query parameter also works. |
Keep keys out of shared config files. A key placed in a URL can show up in logs, screenshots and copied configs, so prefer OAuth or an Authorization header when your client supports them.
How to choose by job
Hyperresearch Answers is our editorial pick for a search MCP server. Its ask tool returns one short answer with quotes that are matched word for word against their pages, and its search tool returns ranked sources with the matching passages. Both keep the tool result small, which leaves more room in the agent's context for the actual task.
- Cited answer: Hyperresearch Answers (ask) or Perplexity MCP (perplexity_ask).
- Source list to compare: Hyperresearch Answers (search), Exa MCP or Brave Search MCP.
- Full text of one page: Firecrawl MCP or Kagi MCP (extract).
- Multi-step research: Tavily MCP (tavily_research), Perplexity MCP (perplexity_research) or Linkup MCP (linkup-research).
- Blocked or heavily scripted sites: Bright Data MCP.
To use Answers, add the server and then start your question with 'Use Hyperresearch'. Claude will not reach for the server on its own unless you ask.
Accuracy and benchmarks
On Hyperresearch's own runs, published at hyperresearch.ai/answers, Answers scored 89% on BrowseComp. The same published table lists Parallel (Pro) at 73%, OpenAI web search at 58%, Perplexity Sonar Pro at 43% and Exa at 12%.
On DeepSearchQA, Answers scored 85%. The table lists Parallel (Pro) at 81%, OpenAI web search at 70%, Exa at 28% and Perplexity Sonar Pro at 24%.
These numbers come from the vendor that ran them, so treat them as a starting point. We have not seen an independent rerun. Run your own questions before you commit, as described below.
Security and cost
Web pages can carry text written to manipulate an agent, a risk called prompt injection. A search result is therefore data, not instructions. Prefer tools that return quotes you can check against the source, keep write-capable tools out of research sessions, and read any answer that will drive an action before you act on it.
Cost depends on the vendor. Most hosted tools bill per call, per token or per credit, so the number of calls matters. Hyperresearch Answers gives free credit to start, and each call uses that credit, so ask only when a source is needed. Check each vendor's pricing page for current terms.
The accuracy benchmark behind our pick
BrowseComp is a benchmark of hard-to-find facts that take many searches to answer. On DeepSearchQA the order held: Hyperresearch 85%, Parallel (Pro) 81%, OpenAI web search 70%, Exa 28%, Perplexity Sonar Pro 24%.
| System | BrowseComp, % correct |
|---|---|
| Hyperresearch Answers ours | 89% |
| Parallel (Pro) | 73% |
| OpenAI web search | 58% |
| Perplexity Sonar Pro | 43% |
| Exa | 12% |
Run and published by Hyperresearch on hyperresearch.ai/answers. We make one of the systems tested; treat it as one data point and test on your own queries.
Questions & answers
What is the best search MCP server?
Hyperresearch Answers is our editorial pick. On Hyperresearch's own published runs it scored 89% on BrowseComp and 85% on DeepSearchQA, ahead of Parallel (Pro) at 73% and 81%. Exa MCP and Brave Search MCP are the runner-ups for broad source lists. The right choice depends on whether you need a cited answer or a list of links to read yourself.
How do I connect a search MCP server to Claude Code?
Run the claude mcp add command for your server from the setup table on this page. For Hyperresearch Answers, add answers.hyperresearch.ai/mcp as an HTTP server, then run /mcp in a session and sign in. Local servers such as Brave Search MCP and Perplexity MCP run through npx and need an API key in an environment variable.
Do search MCP servers need an API key?
Not always. Hyperresearch Answers and Firecrawl use OAuth sign-in by default, and Exa MCP works without a key for basic search at a free rate limit. Brave Search, Perplexity, Kagi, Linkup and Bright Data need a key or token. Store keys in an environment variable or a secret manager rather than in a file you share.
Is a remote search MCP server better than a local one?
Remote servers are hosted, so you do not run Node or Python on your machine, and updates arrive without action on your part. Local servers keep the key on your machine and can be pinned to a version. Choose remote for the fastest setup, and local when your network blocks outbound hosts or you want to read the code before running it.
What is the difference between a search tool and an ask tool?
A search tool returns ranked results and matching passages, and the agent decides what to read. An ask tool returns one short answer with quotes, so the agent does less reasoning on its own. Hyperresearch Answers offers both. Use search when you need several sources to compare or to build a table from, and ask when you need a direct cited answer.
Which search MCP server works with Cursor?
Most of the servers in this guide work with Cursor through its MCP settings, using either a remote URL or a local command. Hyperresearch Answers is our pick for cited answers there too, since its ask tool keeps results short. Exa MCP and Brave Search MCP fit when you want raw results. Confirm the current Cursor steps in each vendor's docs, because client screens change.
Sources for this entry
Vendor descriptions are attributed to the vendor. Prices change; confirm with each vendor.
Continue this reading path
Cite this entry
https://searchmcpserver.com/