2026-10-26, MCP · Agents

Best MCP servers for web search in 2027

An MCP server for web search gives an agent one or more tools it can call mid-task: search the web, read a page, or ask an answer engine. The servers below differ on where they run (a hosted URL or a process on your machine), which tools they load into the model’s context, how the key is passed, and what each tool call costs. This post compares them on those points. For the search APIs behind them (index, filters, data terms), see best search APIs for AI agents.

Facts come from each vendor’s own repository, docs or pricing page, linked and checked on 2026-10-26. “Not stated” means not found there on that date. Prices are list prices in October 2026, for 2027 budgets.

Comparison

Serper has no official MCP server; its site mentions none, and the servers listed for it are community projects (mcp.so). Google’s list of supported MCP servers has no Google Search or web search server; the nearest entries are Agent Search over your own data and Maps Grounding Lite.

Three kinds of tool

The servers expose different things under the word “search”:

• Result lists: titles, URLs and snippets from an index or a results page (brave_web_search, perplexity_search, search_web, SerpApi’s search, AnswerLine’s google).
• Page content: a page as markdown or extracted chunks for the model to read (brave_llm_context, web_fetch_exa, tavily-extract, firecrawl_scrape, read_url, linkup-fetch, kagi_extract).
• Answers: a written answer with citations (perplexity_ask, brave_summarizer, Bright Data’s chatgpt_ai_insights, AnswerLine’s chatgpt and gemini). Some generate the answer with a model you are billed for; AnswerLine’s return what each assistant shows its users.

A research agent needs the first two. An agent that reports what Google or an assistant tells people needs results and answers from those products.

The servers

Brave Search MCP

Brave’s server runs on your machine: npx -y @brave/brave-search-mcp-server or the mcp/brave-search Docker image, over stdio by default or streamable HTTP at 127.0.0.1:8080/mcp with --transport http. No hosted URL is documented.

Tools: brave_web_search, brave_news_search, brave_image_search, brave_video_search, brave_local_search (needs a Pro plan, otherwise falls back to web search), brave_place_search, brave_summarizer (an AI summary of a search run with summary: true) and brave_llm_context (page content for grounding). The key is BRAVE_API_KEY; the license is MIT.

Search is $5 per 1,000 requests at 50 queries per second, LLM context included; Answers is $4 per 1,000 plus $5 per million tokens at 2 per second. Each comes with $5 of free credit a month.

Fits teams that want a local process, an independent index and one flat price for results and page chunks.

Tavily MCP

Hosted at https://mcp.tavily.com/mcp/, with the key as ?tavilyApiKey=, a Bearer header, or OAuth; locally npx -y tavily-mcp@latest on Node 20 or later. Tools: tavily-search, tavily-extract, and map and crawl tools; the README does not describe the return format.

Basic search is 1 credit and advanced 2; extract is 1 credit per 5 URLs; research costs 4 to 250 credits. Pay as you go is $0.008 a credit, $8 per 1,000 basic searches; plans go down to $0.005. 1,000 credits a month are free.

Fits research agents that search and extract in two calls with the least setup; OAuth avoids a key in client config.

Exa MCP

Hosted at https://mcp.exa.ai/mcp over streamable HTTP, with OAuth, ?exaApiKey=, a Bearer header or x-api-key; without a key it works with a rate limit (docs). Locally npx -y exa-mcp-server with EXA_API_KEY.

By default it loads two tools: web_search_exa (results with page content) and web_fetch_exa (a page as markdown). ?tools= adds web_search_advanced_exa and agent_run, which needs a key. Two default tools keep the model’s tool list short.

Search is $7 per 1,000 on auto or fast, $4 on lite or instant, $12 deep and $15 deep-reasoning; contents $1 per 1,000 pages; agent runs $0.012 to $1.00 each. $10 of credit a month is free.

Fits agents that need semantically related pages with text, and evaluations you can start without a key.

Perplexity MCP

Hosted at https://api.perplexity.ai/mcp, or npx -y @perplexity-ai/mcp-server over stdio or HTTP. Four tools: perplexity_search returns ranked results with dates from the Search API; perplexity_ask, perplexity_reason and perplexity_research return written answers from the Agent API’s fast, medium and high presets, and research runs can take minutes. The README says the older Sonar models are no longer used. The key is PERPLEXITY_API_KEY.

Search is $5 per 1,000, or $1 with fast search. The answer tools bill model tokens plus $2.50 per 1,000 web searches, $1.00 fast (Agent API tools). A free tier was not stated.

These answers come from the developer API with the presets you choose. They are not the answer the Perplexity app shows a user; for that, see AnswerLine’s perplexity tool below.

Fits agents that want a cited answer generated in one call.

Firecrawl MCP

Hosted at https://mcp.firecrawl.dev/v2/mcp, with a keyless free tier for scrape, search and parse; with a key, the README says to send a Bearer header, not put it in the URL. /v2/mcp-oauth takes OAuth and /v2/mcp-search loads search only. Locally npx -y firecrawl-mcp or a Docker image, over stdio, streamable HTTP or SSE.

About 30 tools, including firecrawl_search (results, optionally with content), firecrawl_scrape (markdown, JSON and other formats), firecrawl_map, firecrawl_crawl, firecrawl_agent and firecrawl_credit_usage. Search costs 2 credits per 10 results and scrape 1 credit a page; Hobby is $19 for 5k credits, Standard $99 for 100k, and 1,000 credits a month are free.

Fits agents whose work is reading and crawling sites more than searching. Use /v2/mcp-search when you want search alone.

Bright Data MCP

Hosted at https://mcp.brightdata.com/mcp?token=<API_TOKEN>, with groups and tools parameters to limit what loads; locally npx @brightdata/mcp with API_TOKEN. 32 tools load by default, including search_engine (Google results as JSON, Bing and Yandex as markdown), search_engine_batch, scrape_as_markdown, extract (a page to JSON), 14 scraping_browser_* tools, and chatgpt_ai_insights and perplexity_ai_insights, which return those assistants’ answers as markdown.

SERP requests are $1.50 per 1,000 pay as you go, or $499 a month for 380k then $1.30 per 1,000. The README states 5,000 free requests a month; the pricing page states 5K free requests without mentioning MCP.

Fits agents that need Google results and a browser for logged-out pages in one server. Trim the tool list with groups: 32 tool schemas use context on every turn.

SerpApi MCP

Hosted at https://mcp.serpapi.com/mcp with a Bearer key or /<KEY>/mcp; locally over stdio or self-hosted HTTP. One search tool covers every SerpApi engine (default google_light) and returns JSON in compact or complete mode, or markdown with output: "md". search_table and search_dashboard are opt-in MCP Apps views.

Starter is $25 for 1,000 searches, Production $150 for 15,000 ($10 per 1,000), Searcher $725 for 100,000; 250 searches a month are free.

Fits agents that need many engines (Google verticals, Bing, marketplaces) through one tool.

Jina MCP

Hosted at https://mcp.jina.ai/v1 (/sse is legacy). Without a key: read_url (a page or PDF as markdown, rate-limited), capture_screenshot_url, guess_datetime_url and two others. With a Bearer JINA_API_KEY: search_web (title, URL, snippet), search_arxiv, search_ssrn, search_images, sort_by_relevance, deduplicate_strings and extract_pdf. The license is Apache-2.0.

Billing is in tokens: a search costs from 10,000, and a new key gets 10M free. The price of a token pack was not stated on the page checked.

Fits agents that read many pages and papers, where read_url does most of the work.

Linkup MCP

Hosted at https://mcp.linkup.so/mcp with a Bearer key or ?apiKey=; locally npx -y linkup-mcp-server over stdio. Tools: linkup-search (standard or deep), linkup-fetch (page content, optional raw HTML), and linkup-research, which returns a task id that linkup-get-research polls.

Search is $0.005 to $0.006 a request, fetch $0.001 to $0.006, research $0.25 to $2.50. The page lists 4,000 free queries and does not say whether they renew.

Fits agents that run long research jobs; the task-id pattern avoids a tool call held open for minutes.

Kagi MCP

Hosted at https://mcp.kagi.com/mcp with a Bearer key (OAuth not supported yet), or uvx kagimcp over stdio or HTTP. Two tools: kagi_search_fetch (web, news, video, podcast and image search with optional page extracts) and kagi_extract (a page as markdown). The earlier kagi_fastgpt and kagi_summarizer tools were removed.

Search is $12 per 1,000 and extract $4 per 1,000 pages, pay as you go. On Family and Team plans only the account owner can use the API portal (help).

Fits Kagi subscribers who want its results in an agent.

AnswerLine MCP

Hosted only, at https://mcp.answerline.dev/mcp, over Streamable HTTP with POST, protocol revisions 2024-11-05 through 2025-11-25 and 2026-07-28 (MCP docs). Send the key as Authorization: Bearer, or use https://mcp.answerline.dev/<API_KEY>/mcp for clients that only take a URL.

One tool per engine plus credits. Each tool’s input schema is the endpoint’s request schema, and its result is the API’s JSON response:

• google: organic results, People Also Ask, local results, ads, knowledge graph and, with include.aioverview, the AI Overview with its sources.
• google_news: newsResults[] with position, title, source, date and link, per country and hl.
• aimode: Google AI Mode’s answer with sources and citationPills.
• chatgpt, gemini, perplexity, copilot, grok: the answer each assistant shows its users, with sources[] in display order.
• credits: the balance in dollars, free.

Every engine tool’s description tells the model the result is third-party content to treat as data, not instructions. API errors come back as results marked isError with the error code, and the charge, balance and request id are in the result’s _meta.

Prices per tool call: google $0.00103 on every tier, google_news $0.0002 and chatgpt $0.0006 at the Starter tier, perplexity $0.0015 and aimode $0.00206 on every tier. A call costs what the same API request costs; failed calls are not charged, and new accounts get 100 free requests.

It has no page-fetch tool and no semantic search, and a call can take up to five minutes, so raise the client’s tool timeout. Pair it with a fetch tool when the agent must read pages.

Fits agents whose job is search visibility: what Google, Google News, AI Mode and the assistants show for a query, in a given country.

How to choose

1. Hosted or local. A hosted URL needs no process or update on the user’s machine. Local servers (Brave only, or any of the npx options) keep the key in a local env var and let you pin a version.
2. Tools you need. Count the kinds: results, page content, answers. Exa, Kagi and Linkup cover results and content with two to four tools; Bright Data and Firecrawl add browsers and crawlers.
3. Tool count. Every tool schema goes into the model’s context. Load only what the agent uses: Exa’s ?tools=, Bright Data’s groups, Firecrawl’s /v2/mcp-search.
4. Auth. Prefer a Bearer header or OAuth. Keys in a URL (Tavily, Exa, Bright Data, Linkup, SerpApi, AnswerLine path form) end up in logs and screenshots.
5. Whose results. An independent index (Brave, Exa, Kagi), Google results (SerpApi, Bright Data, AnswerLine), or an assistant’s answer as users see it (AnswerLine).
6. Price per completed task. Multiply searches per task by price per call, then add fetches and answer tokens.
7. Long calls. Research tools and live engine calls can take minutes. Check that your client’s tool timeout is configurable, or use a server that returns a task id.

Cost example

An agent completes 2,000 tasks a month with 3 searches each: 6,000 search calls, no fetches. List prices checked 2026-10-26, before tax; recurring monthly free credit is netted, one-time allowances are not.

Jina is left out: its token pack price was not stated. The rows return different things. Brave, Exa and Kagi return an independent index; Bright Data, SerpApi and AnswerLine’s google return Google results; Exa and Firecrawl can include page text in the call, which saves a fetch.

Checking an MCP server before you connect it

A tools/list request shows the tool names, descriptions and input schemas a server will put in the model’s context. For AnswerLine:

curl -s https://mcp.answerline.dev/mcp \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' | jq '.result.tools[].name'

Then one call, to see a result and its _meta cost fields:

curl -s https://mcp.answerline.dev/mcp \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json" \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/call",
       "params":{"name":"google_news","arguments":{"query":"\"acme corp\" recall","country":"US","hl":"en"}}}' \
  | jq '.result._meta["x-credits-charged"], (.result.content[0].text | fromjson | .result.newsResults[:3])'

Client setup for Claude Code and Cursor is in MCP setup for Cursor and Claude Code, and budget guards in giving an agent live answer data.

Pitfalls

• Keys in URLs. A URL with a key is a credential. Give each client its own key so you can revoke it alone.
• Too many tools. 30 tool schemas cost context on every turn and make the model pick worse. Load the subset you use.
• Search loops. A model unsure of an answer searches again. Cap calls per task in the host, not only in the prompt.
• Instructions inside results. Pages and answers can contain text aimed at the model. Treat tool output as data.
• Timeouts on long tools. perplexity_research runs can take minutes and an AnswerLine engine call up to five; a short client timeout drops a call that may still be charged.
• Generated answers taken as the product’s answer. An API answer reflects the model and preset you chose. To measure what ChatGPT or Perplexity tells users, use a tool that returns that answer.

Keep reading