What Is MCP (Model Context Protocol) and Why Search Is a Natural Fit

A plain-language introduction to the Model Context Protocol and why live web search was one of the first tools people wired up with it.

The short version

The Model Context Protocol (MCP) is an open protocol, originally published by Anthropic, for connecting an AI model or agent client to external tools and data sources through a standard interface. Instead of every application inventing its own bespoke way to expose "tools" to a language model, an MCP server declares a set of capabilities — tools, resources, prompts — in a consistent shape, and any MCP-compatible client (Claude Desktop, Claude Code, an IDE, a custom agent loop) can discover and call them the same way.

Why this matters beyond search

Before MCP, wiring a new capability into an agent usually meant writing custom glue code specific to that agent framework: a LangChain tool class here, a different adapter for a different framework there. MCP separates the "I have a capability to expose" problem from the "which agent framework is calling me" problem. A server author writes one MCP server; any conforming client can use it.

Why search is a natural first tool

Search was one of the earliest and most obvious things people built MCP servers for, and it's easy to see why. A language model's knowledge is frozen at training time and has no visibility into anything that happened afterward, or into anything that was never public in the first place. A search tool is the most direct fix: give the model a way to issue a query and get back current, real-world results it can reason over before answering. It's also a relatively simple tool to define — a query in, a list of structured results out — which made it a good early proving ground for the protocol itself.

What a search MCP server actually does

Structurally, a search MCP server is thin: it exposes one or more tools (commonly named something like search or web_search), accepts a query and some optional parameters, calls out to an underlying search API or index, and returns results shaped for a model to consume — typically titles, snippets, and URLs. The MCP server itself usually isn't the search engine; it's a standardized wrapper around one.

Where to go from here

See our Search MCPs hub for a roundup of the notable search MCP servers people have actually built, and how MCP tools differ from plain function-calling if you're deciding whether to adopt the protocol at all.