Web search MCP servers: how to pick one for your agent

A web search MCP server lets an agent search the web the way it calls any other tool. You add one URL to Claude Code, Cursor or Codex, and the agent gets a `search` tool that returns titles, links and snippets. The hard part is choosing one, because MCP server lists now hold dozens of them and they return different things.

This guide sorts web search MCP servers into three kinds, shows what each provider returns through looot's catalog, and gives the live price per search. It ends with one setup that works in Claude Code.

Last checked 2026-10-16, against looot's live catalog and one real search run.

What does a web search MCP server do?

A web search MCP server exposes a search tool over the Model Context Protocol. The agent sends a query, the server calls a search engine or its own index, and returns results as tool output. Most also offer a second tool: read a page, get a direct answer, or crawl a site. The MCP server explainer covers the protocol itself.

Here is a real run from 2026-10-16. We searched "web search mcp server" through looot's `web.search` job. It returned 10 results in about one second, each with a title, URL, snippet and position, and the provider it picked charged nothing for the call.

{
  "query": "web search mcp server",
  "results": [
    { "position": 1, "title": "A simple, locally hosted Web Search MCP server for use with ...", "url": "https://github.com/mrkrsl/web-search-mcp", "site_name": "github.com" }
  ],
  "total_results": 10
}

The three kinds of web search MCP server

Self-hosted servers are free to try. The web-search-mcp project on GitHub, the first result above, says it needs no API keys. A one-provider server is the simplest paid route when you already know which index you want. A multi-provider server makes sense when the agent also needs other data, such as company records or a page as markdown, and you want one key and one bill for all of it.

Which providers can a web search MCP server call?

These providers run `web.search` in looot's catalog today. What each one adds is from the catalog's capability list, not from a test of quality. Each has its own page with the jobs it covers.

Live price per search, read from the catalog when this page loads:

Some endpoints charge per result rather than per call, so a query that asks for 20 results can cost more than one that asks for 5. The web search API pricing post works through what a search costs at volume.

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

The best one is the one your agent needs for its real task, and that is usually one server, not four. Claude Code loads tool definitions on demand with tool search, which is on by default, and warns when one tool's output passes 10,000 tokens (Claude Code MCP docs, read 2026-10-16). Long result lists still cost context, so ask for 5 to 10 results, not 50.

Three questions settle it:

1. Do you want links, or an answer? A plain search returns links. Exa and Linkup also return a direct answer.
2. Will the agent read the pages? Pick a server that can also fetch a page as markdown, or the agent needs a second server.
3. Does it need more than search? Company data, emails or Google rankings come from other providers. One multi-provider server saves you a key per vendor.

To add looot to Claude Code:

claude mcp add --scope user --transport http looot https://api.looot.ai/mcp
claude mcp login looot

The web search API page lists the job's inputs and output fields.

What it cannot do

• looot does not keep its own search index. Results come from the provider that served the run.
• We have not compared result quality across providers. The tables above list what the catalog says each one does.
• It cannot search inside private accounts or behind a login.
• A search that should repeat needs scheduled runs, not available yet.