Best Web Search MCP Servers for AI Agents: How to Choose

Your coding assistant knows a lot, but it does not know what changed in a library last week. A web search MCP server fixes that by letting the assistant query the live web on demand. This guide explains how these servers work, gives five criteria for comparing them, and applies those criteria to the main options.

What Is a Web Search MCP Server?

A web search MCP server is a small program that exposes web search, and often page retrieval, as tools over the Model Context Protocol (MCP). MCP is an open standard that lets AI clients such as Claude Code, Claude Desktop, Cursor, VS Code and Windsurf call outside tools in a consistent way. Once the server is connected, the model can decide to search, read the results and answer from current sources instead of its training data.

There is no single best web search MCP server. The right choice depends on one question: does your agent need links, or does it need the text on the page?

• Quick lookups and links: Brave Search MCP server.
• Agent research with a free monthly allowance: Tavily MCP server.
• Research-style discovery: Exa MCP server.
• Extraction-heavy workflows: Firecrawl MCP server.
• Search plus full page content, batches and crawls in one server: Olostep MCP server.

This page focuses on the MCP layer. For the APIs underneath these servers, see our guide to best web search APIs compared. For the broader architecture, see how to add web search to agents.

How a Web Search MCP Server Works

A web search MCP server sits between your AI client and a search provider. The client lists the server's tools to the model, the model calls a tool with a query, and the server returns results the model can read.

Two design choices shape what your agent gets back: the type of output and how the server runs.

Search Results vs. Page Content

A search tool returns a ranked list of results. Each result usually has a title, a URL and a short snippet of text. Snippets help the model pick a source, but they often do not contain the full answer.

When the snippet is not enough, the agent needs a second step: fetch the page and convert it into text the model can read. This is the search-then-fetch pattern. An Olostep agent, for example, can call a real-time web search API to find sources and then scrape pages into Markdown to read them.

Page fetching has its own condition. JavaScript-heavy websites may need browser rendering before the final text exists. A server that only downloads raw HTML can return an empty or partial page on those sites.

Local Stdio vs. Hosted Servers

MCP defines two standard ways for a client and server to talk. The MCP transports specification describes the local option this way: "stdio: newline-delimited messages over the standard streams of a client-launched subprocess." The second option, Streamable HTTP, sends each message as an HTTP request to a remote endpoint.

In practice, this gives you two setups:

• Local stdio: The client starts the server on your machine, usually with npx. This option suits offline work, corporate proxies and air-gapped setups. It requires a local runtime such as Node.js.
• Hosted (Streamable HTTP): You paste a URL and an auth header into your client. Nothing is installed locally, and the provider handles updates.

Five Criteria for Choosing a Web Search MCP Server

Search quality matters, but it is only one part of the decision. These five criteria cover what changes your agent's results and your costs in daily use. The comparison later in this article draws on them.

Output Shape: Snippets, Full Text or Answers

Web search MCP servers return one of three output shapes:

• Snippets: Titles, URLs and short excerpts. These are cheap and fast, and they work well when the agent only needs a link or a quick fact.
• Full page content: The page converted to Markdown or text. The agent needs this when it must read documentation, compare details or quote a source.
• Synthesized answers: A written answer with source links. This is useful when you want one tool call per question, but the server's model shapes the answer, not yours.

Olostep's Answers endpoint is an example of the third type. It returns grounded answers with citations and can format the answer to a JSON schema you provide.

Tool Footprint and Context Cost

Every tool a server exposes has a definition: a name, a description and an input schema. The client loads these definitions into the model's context before the conversation starts. More tools give the model more options, but they also use more of the context window.

The cost adds up across servers. In one example with five connected servers, Anthropic's tool use research reports: "That's 58 tools consuming approximately 55K tokens before the conversation even starts."

To keep the footprint small:

• Enable only the servers you use: Disconnect servers you don't need for the current project.
• Prefer focused toolsets: A one-tool search server costs less context than a ten-tool server. It also does less.
• Use client-side tool search where available: Some clients can load tool definitions on demand instead of all at once.

Pricing Unit and Free Tier

Vendors bill in different units. Some charge per request, some charge per credit, and some give a monthly dollar credit. A "free tier" can mean a fixed number of searches or a dollar amount that covers different numbers of searches depending on the search type.

Compare prices in each vendor's own unit, and note the date you checked. Free tiers in this category change often.

Setup, Auth and Client Support

Servers authenticate in three common ways:

• Environment variable: Local stdio servers usually read an API key from an environment variable.
• Bearer header: Hosted servers usually expect Authorization: Bearer YOUR_API_KEY in the client config.
• No key: Some free servers need no key because they scrape public search pages.

Check that the server documents setup for your client. Config formats differ: VS Code, Windsurf and Claude Code each use slightly different keys. If your team shares a project-level MCP config file, keep API keys out of it and load them from the environment instead.

Safety of Retrieved Content

Every page a search server returns is untrusted input. A page can contain text written to manipulate a model. OWASP prompt injection guidance defines the risk directly: "Indirect prompt injections occur when an LLM accepts input from external sources, such as websites or files."

A search server cannot fully remove this risk, so build guardrails around it:

• Keep approval on side effects: Require human confirmation before the agent runs tools that write files, send messages or spend money.
• Separate reading from acting: Treat fetched text as data to analyze, not as instructions to follow.
• Limit scope: Give agents that browse the web the smallest set of other tools they need.

The Best Web Search MCP Servers Compared

Olostep publishes this article and is one of the options below. Every server is compared on the same table columns, and pricing is quoted in each vendor's own unit.

Pricing, checked on October 2, 2026:

Sources:

• Brave: Brave Search API pricing states: "$5.00 per 1,000 requests. Includes free $5 in credits every month."
• Tavily: Tavily credits and pricing states: "You get 1,000 free API Credits every month. No credit card required."
• Exa: Exa API pricing states: "The Free Tier gives you $10 in credits (up to 2,500 Instant searches) the day you sign up, and your free balance resets to $10 on the first of every month."
• Olostep: Olostep pricing plans list a free trial with 500 successful requests and a $9/month Starter plan with 5,000 successful requests.

Olostep MCP Server: Search Plus Page Retrieval in One Server

The Olostep MCP server connects MCP clients to Olostep's web search, scraping and URL discovery. According to the Olostep MCP documentation, the full server exposes 10 tools for the live web. The search tool and the page tools work together, so one server covers the full search-then-fetch pattern.

• search_web: Returns structured JSON search results rather than AI-written prose.
• scrape_website and get_webpage_content: Return a page as Markdown, HTML, JSON or text, with optional JavaScript rendering.
• batch_scrape_urls: Scrapes 2 to 10,000 known URLs asynchronously and returns a batch ID for collecting results.
• create_crawl and create_map: Follow links from a start URL or list the URLs on a site.
• answers: Returns a sourced answer, optionally shaped to a JSON schema.

Olostep offers a hosted endpoint at https://mcp.olostep.com/mcp and a local stdio install through npx. The trade-off is footprint: 10 tool definitions use more context than a one-tool search server. If you only need links, a smaller server may be the better fit.

Best for: Agents that need to read documentation, scrape JavaScript-heavy pages or process many URLs from a single search.

Brave Search MCP Server: Fast Results From the Brave Search API

Brave maintains an official Brave Search MCP server that connects clients to the Brave Search API. It returns search results with snippets, which works well when the agent needs links or quick facts.

If your agent needs full page text, pair it with a fetch or scrape tool. Billing is per request, with a monthly credit covering light use.

Best for: Fast lookups, link discovery and pipelines that already have a page fetcher.

Tavily MCP Server: Search Built for LLM Agents

Tavily maintains the Tavily MCP server, which exposes search and content extraction tools. Results are formatted for language models rather than for human readers.

Pricing is credit based, so check how many credits each search depth uses before you scale up.

Best for: Prototyping research agents within a monthly free allowance.

Exa MCP Server: Search for Research Queries

Exa maintains the Exa MCP server for its search API. It suits research questions and "find pages like this" tasks.

Exa's price per search depends on the search type. The same free credit covers a different number of searches depending on which type you call.

Best for: Research-style queries and finding similar pages.

Firecrawl MCP Server: Search and Extraction

Firecrawl maintains the Firecrawl MCP server, which combines search with scraping and extraction tools. In the AIMultiple MCP benchmark, updated March 2026: "Firecrawl is the fastest MCP with the average MCP run time for correct results of 7 seconds and its accuracy rate was 83%."

AIMultiple's tasks centered on shopping and LinkedIn navigation, not documentation lookups. Servers that scored low there may perform differently on documentation or news queries. Run your own test on the queries your agent actually handles.

Best for: Extraction-heavy workflows that turn pages into structured data.

Free and Self-Hosted Options: DuckDuckGo and Google Scrapers

Several open-source MCP servers search DuckDuckGo or Google without an API key. They work by scraping public search result pages.

Free scrapers have known limits:

• Rate limits and blocks: Search engines may throttle or block automated requests, especially at volume.
• Breakage: When the search page's markup changes, the scraper can stop returning results until it is updated.
• Terms of service: Check the search engine's terms before relying on scraped results in a product.

Best for: Personal experiments and learning how MCP tools work.

How to Add a Web Search MCP Server to Claude Code or Cursor

The steps below use the Olostep hosted endpoint. Most hosted servers follow the same pattern with a different URL and key.

Prerequisites: An Olostep API key from the dashboard and an MCP-compatible client.

1. Copy your API key from the Olostep dashboard.
2. Add the server to your client. In Claude Code, run:

claude mcp add --transport http olostep https://mcp.olostep.com/mcp \
  --header "Authorization: Bearer YOUR_API_KEY"

In Cursor or Claude Desktop, add this to your MCP config file:

{
  "mcpServers": {
    "olostep": {
      "url": "https://mcp.olostep.com/mcp",
      "headers": { "Authorization": "Bearer YOUR_API_KEY" }
    }
  }
}

1. Restart the client so it loads the new server and its tools.
2. Verify the connection. Ask a question that needs current information, such as "Search for the latest release notes for Next.js and summarize them." The client should show a search_web call followed by a page retrieval call.

Local option: To run the server on your machine instead, use npx -y olostep-mcp as the command and set OLOSTEP_API_KEY as an environment variable. This requires Node.js 18 or later.

Troubleshooting: A common setup error is mixing auth modes. The hosted endpoint expects the Bearer header, while the local stdio server expects the environment variable.

FAQ

Do I Need an API Key for a Web Search MCP Server?

Most hosted web search MCP servers need an API key to track usage and billing. Keyless servers exist, but they scrape public search pages and can break or get blocked.

Is There a Free Web Search MCP Server?

Yes, Brave, Tavily, Exa and Olostep all offer free monthly credits or free starter requests. Self-hosted DuckDuckGo and Google scrapers cost nothing but are less reliable.

Can I Use More Than One Search MCP Server at Once?

Yes, you can connect several servers, and the model will choose between their tools. Each server adds its tool definitions to the context window, so connect only the ones you use.

Does a Web Search MCP Server Read JavaScript-Heavy Pages?

Only if the server renders pages before extracting text. Search-only servers return snippets and need a separate fetch tool that supports JavaScript rendering.

Should I Use My Client's Built-In Web Search Instead?

Built-in search is the simplest option when it covers your needs. An MCP server gives you control over the search provider, the output format and whether the agent can read full pages.