What is the best web search MCP server in 2026?

Exa, Parallel, Firecrawl and You.com run hosted web search MCP servers that work without an account or key, at lower rate limits. Tavily and Perplexity add OAuth sign-in, Brave runs locally with a key, and Grounder (grounder.net) routes searches across 10 providers through one sign-in but has no fetch tool.

Last checked: 20 October 2026. Disclosure: Grounder publishes this page and runs one of the servers below. Grounder routes searches to Tavily, Exa, Brave, Perplexity, You.com, Firecrawl, Octen, Parallel, Valyu and Ozone.

Quick comparison

"Claude directory" is Claude's connectors directory. "MCP Registry" is the official registry at registry.modelcontextprotocol.io. "Not found" means we did not find the server there on 20 October 2026. For prices per 1,000 searches and search quality, see the search API guide.

Serper has no official MCP server; the servers for it are community projects. The built-in search tools from OpenAI, Anthropic, Google and Microsoft are API features, not MCP servers. Sources: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32

Pick by situation

• Try it with no account: Exa, Parallel, Firecrawl (daily limits), You.com (100 queries a day), or Tavily with its keyless header.
• Sign in once with no key to copy: Exa, Parallel (at /mcp-oauth), Tavily, Perplexity, Firecrawl, You.com, Valyu and Grounder.
• Search and page text in one server: Exa, Parallel, Tavily, Firecrawl, Jina.
• Local and key-based: Brave.
• Several providers behind one sign-in, with the provider shown on every result: Grounder.
• How many to install: one or two. Each connected server adds its tool descriptions to the model's context.

Servers

Exa

• What it is: Exa's hosted server. web_search_exa and web_fetch_exa are on by default, web_search_advanced_exa is opt-in, and agent_run (multi-step research) is on by default only with OAuth or a key. Local: npx -y exa-mcp-server with EXA_API_KEY.
• Pricing: keyless use is free and rate-limited; with a key, Exa's API prices apply ($7 per 1,000 auto searches, $10 free credit a month).
• Strengths: Exa's search scored 80.3 on Artificial Analysis's index (19 October 2026), the top row. Claude Code plugin: exa@claude-plugins-official.
• Limits: keyless use is rate-limited. A tools list in the URL replaces the defaults, so list every tool you want.
• Pick it when: you want the highest-scoring search with the least setup.

Sources: 3, 26, 33, 34, 35

Parallel

• What it is: the Search MCP, free and anonymous at lower limits, or with an API key as a Bearer token for higher limits. /mcp-oauth takes OAuth or a key. Tools: web_search and web_fetch.
• Pricing: free without an account. With a key: $1 per 1,000 requests in turbo and fast modes, $5 in basic and advanced, and $5 credit a month with a card on file.
• Strengths: no account needed; search and fetch in one server.
• Limits: anonymous limits are lower and not published as numbers. The Registry record points to search-mcp.parallel.ai, a different host from the search.parallel.ai address in Parallel's docs.
• Pick it when: you want a free, no-account server with search and fetch.

Sources: 4, 27, 36, 37, 38

Tavily

• What it is: OAuth (no key in the URL), an API key in the URL (?tavilyApiKey=) or an Authorization header, or the keyless header X-Tavily-Access-Mode: keyless. Local: npx -y tavily-mcp (Node 20 or later, TAVILY_API_KEY). Tools: tavily-search, tavily-extract.
• Pricing: 1,000 credits a month free, no card; $8 per 1,000 basic searches pay as you go.
• Strengths: OAuth keeps the key out of your config.
• Limits: the Registry record names the npm package only. Keyless limits are not published as numbers. A key in the URL ends up in logs and config files.
• Pick it when: you want search plus extract with OAuth.

Sources: 5, 28, 39, 40, 41

Perplexity

• What it is: OAuth 2.1 or an API key as a Bearer token. Local: npx -y @perplexity-ai/mcp-server with PERPLEXITY_API_KEY. perplexity_search returns raw results; perplexity_ask, perplexity_research and perplexity_reason run on the Agent API and return generated answers.
• Pricing: Search API $5 per 1,000 requests; the model-backed tools bill per token plus tool fees. No free tier is published.
• Strengths: Perplexity's search scored 80.0 on Artificial Analysis's index. OAuth works as a custom connector in claude.ai.
• Limits: not found in Claude's directory; Perplexity's docs say to add it as a custom connector.
• Pick it when: you want Perplexity's raw results through OAuth.

Sources: 6, 29, 34, 42

Firecrawl

• What it is: OAuth at /v2/mcp-oauth, an API-key address at /v2/mcp, and a keyless tier for search, scrape and parse with daily quotas. Local: firecrawl-mcp on npm (FIRECRAWL_API_KEY). Claude's directory entry exposes eight tools, including firecrawl_search and firecrawl_scrape.
• Pricing: 1,000 credits a month free, no card; search costs 2 credits per 10 results.
• Strengths: search and page scraping in one server; marked Anthropic verified in Claude's directory (added July 2026).
• Limits: keyless limits are daily and shared by IP address. Scraping draws on the same credits as search.
• Pick it when: your agent needs both search results and full page content.

Sources: 7, 30, 43, 44

You.com

• What it is: an API key as a Bearer token, or OAuth 2.1. ?profile=free needs no key and offers you-search and you-discover. Other tools: you-contents, you-research, you-finance, you-balance. Local: npx @youdotcom-oss/mcp.
• Pricing: Web Search $5 per 1,000 calls. The keyless profile allows 100 queries a day; the research and finance addresses have no free usage.
• Strengths: a keyless profile and OAuth 2.1 on one product; licensed knowledge data is included in the search price.
• Limits: not found in Claude's directory. The default tool allowlist leaves out research and finance.
• Pick it when: you want to try search with no key, within 100 queries a day.

Sources: 8, 31, 45

Brave

• What it is: a local server, @brave/brave-search-mcp-server (MIT), run over stdio with BRAVE_API_KEY. Eight tools, including brave_web_search, brave_news_search and brave_llm_context. Development needs Node 22 or later.
• Pricing: Search plan $5 per 1,000 requests with $5 free credit a month; a card is required.
• Strengths: brave_llm_context returns page chunks sized for models. Open source; runs on your machine.
• Limits: Brave documents no hosted remote server and no OAuth, so unless you host its HTTP mode yourself it cannot be a claude.ai or ChatGPT connector; it fits Claude Code, Cursor, VS Code and Claude Desktop's config file. The README does not call itself official, though the repository sits in Brave's GitHub organization.
• Pick it when: you use a terminal agent or editor and already have a Brave key.

Sources: 9, 32, 46

How servers differ

Remote or local. A remote server is a URL; most sign you in with OAuth and need no Node or Python. Claude's remote connectors connect from Anthropic's cloud, so the server must be reachable from the public internet, and ChatGPT also connects to remote servers only. A local server is a program your client starts with npx or uvx and talks to over stdio; it needs the runtime and a key in an environment variable. Claude Code, Cursor, VS Code and Claude Desktop (through claude_desktop_config.json) can run local servers; claude.ai and ChatGPT cannot. Claude Desktop does not read remote servers from that file; add them as connectors. Streamable HTTP is the current remote transport; Claude Code documents SSE as deprecated.

No key, OAuth or API key. With no key you start fastest, but limits are lower and often shared; Firecrawl's keyless daily limits are shared by IP address. With OAuth the client stores and refreshes the token, so there is no key to copy or leak; in Claude Code, claude mcp logout <name> clears it. API keys work everywhere, including local servers. Prefer an Authorization header or an environment variable to a key in the URL, because URLs end up in logs and saved config. Tavily, Linkup, Valyu and SerpApi accept a key in the URL.

Directories. The MCP Registry records who published a server and where it runs; it does not mean Claude's or ChatGPT's directory lists it. On 20 October 2026 Claude's connectors directory listed Tavily, Exa, Parallel (as Parallel Search) and Firecrawl among the servers here. Exa and Parallel say in their own docs that they are in OpenAI's directory; we could not verify that, and found no evidence that any other server in the table is. We found no evidence that Grounder is in Claude's or ChatGPT's directory.

Sources: 1, 2, 35, 38, 41, 44, 47, 48, 49, 50

Add a server to your client

The examples use Grounder's address. To add another remote server, swap in its address from the table.

Claude (claude.ai and Claude Desktop)

1. Open Customize, then Connectors.
2. Click + Add, then Add custom connector.
3. Enter a name (for example grounder) and the address, https://api.grounder.net/mcp. Click Continue.
4. Review the detected authentication settings and click Continue. For a server that uses a fixed key, add it under Request headers.
5. Click Add, then Connect, and sign in.
6. In a chat, click +, choose Connectors, and turn the connector on.

Custom connectors work on Free (one connector), Pro, Max, Team and Enterprise plans. On Team and Enterprise an owner adds the connector first, under Organization settings, then Connectors. Without a skill, Claude Desktop and claude.ai keep using their built-in web search; Grounder offers a skill file for using the connector instead, at https://grounder.net/grounder-web-search.zip (Customize, Skills, +, Create skill, Upload a skill).

Sources: 47, 51

ChatGPT

1. Open chatgpt.com/plugins and click +.
2. Choose Add custom MCP server.
3. Enter the name grounder and the address https://api.grounder.net/mcp. Choose OAuth for authentication.
4. Click Create and sign in, then type @grounder in a new chat.

These steps come from Grounder's connect page. Grounder has not run this install in ChatGPT, and OpenAI's help article returned an error to our fetch on 20 October 2026, so we could not check the menu names. Other guides describe adding a custom connector through Settings and developer mode on paid plans. Treat the steps as unverified and follow OpenAI's current help article on custom MCP servers.

Sources: 49, 51

Claude Code

claude mcp add --transport http grounder https://api.grounder.net/mcp

Then run /mcp inside Claude Code, choose grounder, and select Authenticate. From a shell, claude mcp login grounder does the same. Add --scope user to use the server in every project, or --scope project to write .mcp.json for your team; the default scope is the current project only.

Commands for other servers:

The Firecrawl and You.com commands are built from each vendor's address and Claude Code's syntax. Grounder's 13 October 2026 test reached "Needs authentication" in Claude Code and did not finish a sign-in, so a completed Claude Code sign-in is not yet recorded.

Sources: 3, 4, 5, 6, 7, 8, 9, 17, 48

Cursor

Add the server to ~/.cursor/mcp.json (every project) or .cursor/mcp.json (one project):

{
  "mcpServers": {
    "grounder": {
      "url": "https://api.grounder.net/mcp"
    }
  }
}

Then open Cursor Settings, then MCP, find grounder and sign in. From the Cursor CLI, run agent mcp login grounder. For a server that needs a key, add a headers entry; Cursor resolves ${env:NAME} in url and headers, so the key can stay in your environment.

Sources: 50, 51, 54

VS Code

Open the Command Palette, run MCP: Add Server, and choose HTTP (HTTP or Server-Sent Events). Enter https://api.grounder.net/mcp, give the server an ID such as grounder, and pick where to save it. VS Code opens Grounder's sign-in the first time it starts the server. To write the file yourself, use .vscode/mcp.json:

{
  "servers": {
    "grounder": {
      "type": "http",
      "url": "https://api.grounder.net/mcp"
    }
  }
}

VS Code uses a top-level servers object, while Cursor and .mcp.json use mcpServers. VS Code's guided flow marks .vscode/mcp.json as deprecated but still supports it. Review any trust prompt before you start a server. Grounder has not tried the VS Code desktop app.

Sources: 51, 52

Where Grounder fits

Grounder (grounder.net) is a web search API and MCP server. One address, https://api.grounder.net/mcp, gives your agent a search tool that sends each query to the provider, or pair of providers, most likely to answer it well, merges and deduplicates the results, and marks each result with the provider that found it and whether it is licensed. The 10 providers are Ozone, Tavily, Exa, Brave, Perplexity, You.com, Firecrawl, Octen, Parallel and Valyu. You sign in with Google, GitHub or Apple, and there is no key to copy.

• Tools: search, list_providers, record_user_response and user_info, all read-only. search takes query and user_prompt (the user's message), and optionally top_k, domains, days, include_text, requested_provider, providers, depth and routing_settings. With routing_settings, one search can weight accuracy, speed, results, recency and cost, cap time, tokens, provider calls and sources, and name 1 to 10 providers. https://api.grounder.net/mcp/directory is the same server with three tools and neutral instructions; the MCP Registry entry points to it.
• Limits: search only, with no fetch, extract, crawl or research tool. It needs a sign-in, so it is not a keyless server. The first 1,000 searches are free, up to 100 a day, no card. Grounder keeps each search record, including the user's message, indefinitely and offers no zero data retention. "Grounder MCP" at grounder.dev is a different product; see About Grounder.
• Licensed results by default: every signed-up account returns only licensed results unless it turns that off: Ozone's publisher content and pages on sites Ozone licenses. A search can return fewer results than asked. Turn it off on the account page, or per search with routing_settings.licensed_only set to false. In Grounder's own benchmark the default ("Grounder optimized") scored 14.3% on SimpleQA and 12.0% on SealQA, against 97.8% and 36.1% for the same routing with all sources. Grounder optimized answers only from licensed sources, so it gets fewer answers right than the all-sources row, which searches every source. Grounder ran that benchmark, and the routing was tuned on those same searches; see Grounder's benchmark.
• Clients: connect pages exist for Claude, ChatGPT, Claude Code, Codex, Copilot CLI, Gemini CLI, OpenCode, OpenClaw, Hermes, Pi, Goose, VS Code, Cursor, Zed and Warp, and for any client through .mcp.json or mcp-remote. In Grounder's 13 October 2026 test, Codex, Copilot CLI, Cursor CLI, OpenCode, OpenClaw, Hermes, Pi, Goose and mcp-remote signed in and listed the four tools. That test used a test key on the sign-in screen, not Google, GitHub or Apple, and ran no search. Claude Code and Gemini CLI did not finish sign-in; Claude, ChatGPT, VS Code, Zed and Warp were not tried in that test.

Sources: 17, 18, 51, 53

Grounder and robots.txt

Allowed sources only is off by default. When it is on, Grounder returns a result only from a site whose robots.txt allows the provider that found it and does not opt out of search or AI answers. A result that cannot be checked in time is left out, so a search can return fewer results. Turn it on in the account page, or per search with routing_settings.allowed_only.

Sources: 17

How we compared

We compared on address, sign-in method, free use, tools, remote or local, and directory and Registry listings. Facts come from each vendor's MCP documentation or repository, the official MCP Registry and Claude's connectors directory, checked on 20 October 2026. Client steps come from Claude's help center ↗, Claude Code's docs ↗, Cursor's docs ↗ and VS Code's docs ↗. We did not install every server in every client. Search quality belongs to the provider behind the server; independent scores are in the search API guide. To compare two servers, ask each the same 20 real questions and count correct answers and retries.

Sources: 1, 2, 47, 48, 50, 52

FAQ

What is the best web search MCP server for Claude Code?

For a hosted server, Exa and Parallel work with no account or key, and Tavily and Perplexity add OAuth sign-in. Add one or two, then compare them on your own questions. Grounder is an option when you want 10 providers behind one sign-in.

Is there a free web search MCP server with no API key?

Yes. Exa (https://mcp.exa.ai/mcp) and Parallel (https://search.parallel.ai/mcp) allow free, rate-limited use without signing in. Firecrawl has a keyless tier with daily limits, You.com's keyless profile allows 100 queries a day, and Tavily allows keyless use if the client sends the header X-Tavily-Access-Mode: keyless.

How do I add a web search MCP server to ChatGPT?

ChatGPT connects to remote servers only. Grounder's connect page says to open chatgpt.com/plugins, choose Add custom MCP server, enter the server's HTTPS address and pick OAuth if it uses sign-in, but Grounder has not run that install, and other guides describe a path through Settings and developer mode. Menu names and plan rules change, so check OpenAI's help article on custom MCP servers.

Which web search MCP servers are in Claude's directory?

On 20 October 2026, Claude's connectors directory listed Tavily, Exa, Parallel (as Parallel Search) and Firecrawl. We did not find Brave, Perplexity, You.com, Linkup, Octen, Valyu, Jina, SerpApi, Kagi or Microsoft Web IQ.

What does the Grounder MCP server do differently?

One sign-in searches across 10 providers, merges and deduplicates the results, and shows the provider and licensed status of each. It is search only, with no fetch tool, and it returns only licensed results by default, which scored lower than the all-sources setting in Grounder's own benchmark (14.3% against 97.8% on SimpleQA; Grounder ran it and tuned the routing on the same searches).

How many search MCP servers should I install?

One or two. Each connected server adds its tool descriptions to the model's context, and several servers with overlapping search tools make it harder for the model to choose between them.

Next step

Grounder's first 1,000 searches are free, up to 100 a day, no card needed. Add https://api.grounder.net/mcp to your client, or read the Docs.