Web Search MCP Servers Compared (2026): Tavily, Exa, Brave Search, Perplexity, Parallel, SerpApi and Firecrawl

• Tavily: search results and page extraction (tavily-search, tavily-extract). Free: 1,000 credits/month. Paid: $0.008 per credit pay as you go; a basic search is 1 credit, so about $8 per 1,000 basic searches. Cheapest plan: Project, $30/month for 4,000 credits.
• Exa: search results with clean page content (web_search_exa) and full-page fetch (web_fetch_exa). Free: $10 in credits, reset monthly (up to 2,500 Lite searches). Paid: from $4 per 1,000 searches (lite), $7 per 1,000 for auto. Up to 10 results included.
• Brave Search: results from Brave's own index, plus news, images, video, local and LLM context tools. Free: $5 in credits every month. Paid: $5 per 1,000 requests.
• Perplexity: raw search results (perplexity_search) or written answers with web search (perplexity_ask, perplexity_reason, perplexity_research). Search API: $5 per 1,000 requests, or $1 per 1,000 with Fast Search. No free tier is stated on the pricing page.
• Parallel Search: web_search returns results with excerpts; web_fetch returns Markdown from URLs. Free: the hosted MCP server works with no API key at lower rate limits. API: $1 per 1,000 turbo or fast searches, $5 per 1,000 basic or advanced, 10 results included.
• SerpApi: search results from Google (default engine google_light), Bing, Yahoo, DuckDuckGo, YouTube, eBay and more through one search tool. Free: 250 searches/month. Cheapest plan: Starter, $25/month for 1,000 searches (about $25 per 1,000).
• Firecrawl search: firecrawl_search returns results and can scrape each result's page. Free: 1,000 credits/month. Search costs 2 credits per 10 results. Cheapest plan: Hobby, $19/month for 5,000 credits.

Tavily's server gives the agent search and page extraction built for LLMs. The hosted endpoint takes your key in the URL or signs you in with OAuth. The local server reads TAVILY_API_KEY.

claude mcp add --transport http tavily https://mcp.tavily.com/mcp

{
  "mcpServers": {
    "tavily-mcp": {
      "command": "npx",
      "args": [
        "-y",
        "tavily-mcp@latest"
      ],
      "env": {
        "TAVILY_API_KEY": "your-api-key-here"
      }
    }
  }
}

• Hosted with a key: https://mcp.tavily.com/mcp/?tavilyApiKey=<your-api-key>.
• Main tools: tavily-search and tavily-extract. The server also has map and crawl tools.
• Optional env vars: DEFAULT_PARAMETERS sets default values for tavily-search (for example include_images); TAVILY_HUMAN_ID is sent as the X-Human-Id header.
• Gotcha: advanced search costs 2 credits instead of 1, so setting search_depth to advanced as a default doubles the cost.

• Tavily MCP server: install for every agent

Exa returns search results with clean page text, so the agent often does not need a second fetch. The hosted server works without a key, with rate limits. Sign in with OAuth or add an API key for higher limits and for agent_run.

claude plugin install exa@claude-plugins-official

{
  "mcpServers": {
    "exa": {
      "command": "npx",
      "args": [
        "-y",
        "exa-mcp-server"
      ],
      "env": {
        "EXA_API_KEY": "your_api_key"
      }
    }
  }
}

• Hosted: https://mcp.exa.ai/mcp. Pass a key as ?exaApiKey=, an Authorization: Bearer header or an x-api-key header.
• Default tools: web_search_exa and web_fetch_exa. agent_run (multi-step research and structured output) is on with OAuth or an API key. web_search_advanced_exa (domain, date and other filters) is opt-in.
• Gotcha: adding tools to the URL replaces the defaults, so list every tool you want.

• Exa MCP server: install for every agent

Brave's server searches Brave's own independent index. Brave says queries end on infrastructure it runs end to end, with no third party to keep them. There is no hosted MCP endpoint in the README; run it locally with npx or Docker.

{
  "mcpServers": {
    "brave-search": {
      "command": "npx",
      "args": [
        "-y",
        "@brave/brave-search-mcp-server",
        "--transport",
        "stdio"
      ],
      "env": {
        "BRAVE_API_KEY": "YOUR_API_KEY"
      }
    }
  }
}

• Docker: docker run -i --rm -e BRAVE_API_KEY docker.io/mcp/brave-search.
• Main tools: brave_web_search, brave_llm_context (pre-extracted page content for grounding), brave_news_search, brave_image_search, brave_video_search, brave_local_search, brave_place_search, brave_summarizer.
• Env vars: BRAVE_API_KEY (or BRAVE_API_KEY_FILE); BRAVE_MCP_ENABLED_TOOLS and BRAVE_MCP_DISABLED_TOOLS limit which tools load.
• Gotcha: the server defaults to stdio. Use --transport http only when you run it as an HTTP server.

• Brave Search MCP server: install for every agent

Perplexity's official server has four tools. perplexity_search returns raw results from the Search API. The other three return written answers: perplexity_ask uses the Agent API fast preset, perplexity_reason uses medium and perplexity_research uses high. Earlier versions used the Sonar models sonar-pro, sonar-reasoning-pro and sonar-deep-research.

claude mcp add perplexity --env PERPLEXITY_API_KEY="your_key_here" -- npx -y @perplexity-ai/mcp-server

claude mcp add --transport http perplexity https://api.perplexity.ai/mcp --header "Authorization: Bearer YOUR_API_KEY"

• Main tools: perplexity_search, perplexity_ask, perplexity_reason, perplexity_research.
• Env vars: PERPLEXITY_API_KEY (required), PERPLEXITY_TIMEOUT_MS (default 5 minutes), PERPLEXITY_BASE_URL.
• Gotcha: research runs can take minutes. Raise PERPLEXITY_TIMEOUT_MS for long queries.
• Gotcha: the $5 per 1,000 Search API price covers perplexity_search only. The answer tools are billed through the Agent API.

• Perplexity MCP server: install for every agent

Parallel's Search MCP is free to use with no API key, at lower rate limits. Add a key through the OAuth-capable endpoint for heavier use.

claude mcp add --transport http "Parallel-Search-MCP" https://search.parallel.ai/mcp

• With a key: https://search.parallel.ai/mcp-oauth with Authorization: Bearer <PARALLEL_API_KEY>, or the OAuth sign-in.
• Main tools: web_search and web_fetch (Markdown from specific URLs).
• Env var used in the docs: PARALLEL_API_KEY (key from platform.parallel.ai).
• API pricing: $1 per 1,000 turbo or fast requests, $5 per 1,000 basic or advanced, 10 results included.

• Parallel Search MCP server: install for every agent

SerpApi returns structured results scraped from Google and other engines. One search tool covers every SerpApi engine. The hosted endpoint takes your key in the URL path.

claude mcp add --transport http serpapi https://mcp.serpapi.com/YOUR_SERPAPI_API_KEY/mcp

• Header option: send Authorization: Bearer YOUR_API_KEY to https://mcp.serpapi.com/mcp.
• Main tool: search. Opt-in tools search_table and search_dashboard render results in MCP Apps hosts.
• The default engine is google_light. Engines include Google, Bing, Yahoo, DuckDuckGo, YouTube and eBay; the list is at the serpapi://engines resource.
• Self-host: clone the repo, then uv sync && uv run src/server.py.
• Gotcha: at about $25 per 1,000 searches on Starter, it is the most expensive per query here. Use it when you need Google's own result pages.

• SerpApi MCP server: install for every agent

firecrawl_search searches the web and can scrape each result into page content in the same call. Set scrapeOptions to control that. Firecrawl also runs a search-only hosted endpoint with firecrawl_search and five firecrawl_research_* tools and no page fetching.

{
  "mcpServers": {
    "firecrawl": {
      "command": "npx",
      "args": [
        "-y",
        "firecrawl-mcp"
      ],
      "env": {
        "FIRECRAWL_API_KEY": "fc-YOUR_API_KEY"
      }
    }
  }
}

• Hosted: https://mcp.firecrawl.dev/v2/mcp with Authorization: Bearer <FIRECRAWL_API_KEY>, or https://mcp.firecrawl.dev/v2/mcp-oauth to sign in. Never put the key in the URL.
• Search-only endpoint: https://mcp.firecrawl.dev/v2/mcp-search. It needs authentication on every request.
• firecrawl_search parameters include query, limit, lang, country, scrapeOptions, highlights and categories.
• Gotcha: a search costs 2 credits per 10 results. Passing the result id to firecrawl_search_feedback refunds 1 credit, up to a daily cap (default 100 credits per team).
• Self-hosted Firecrawl: set FIRECRAWL_API_URL.

• Firecrawl MCP server: install for every agent

• Cheapest: Parallel Search (free hosted MCP with no key; $1 per 1,000 fast API searches). Perplexity Fast Search is also $1 per 1,000.
• Research and deep answers: Perplexity perplexity_research, or Exa agent_run.
• Google results: SerpApi.
• Privacy and an independent index: Brave Search.
• Full page content with results: Firecrawl search with scrapeOptions, Exa web_search_exa, or Tavily with tavily-extract.
• Keep API keys out of committed files: use environment variables, headers or the OAuth endpoints where the vendor offers them.
• Prices and free tiers were checked on each vendor's pricing or docs page on 2026-10-23 and change often.

• Set up MCP servers, rules and permissions in one paste
• Add an MCP server in every AI coding agent
• Web scraping MCP servers compared