What Is an MCP Server?

The Model Context Protocol (MCP) is a standard for connecting AI agents to external tools and data. An MCP server exposes a set of tools — functions the AI agent can call — with descriptions of what they do and what parameters they accept. The AI agent reads the tool descriptions, decides when to use them, and calls them as needed during a conversation.

If you've ever wished an AI assistant could look something up instead of guessing, MCP is the infrastructure that makes that possible.

Why MCP exists

Large language models are trained on static data. They learn facts from training corpora that are months or years old by the time you talk to them. Ask Claude or ChatGPT for the horsepower of a specific vehicle trim and it might be right — or it might confidently give you a number from two model years ago because that's what appeared in its training data.

This is the hallucination problem, and it's not a bug in the model. It's a fundamental limitation of systems that generate answers from learned patterns rather than querying authoritative sources at runtime.

MCP solves this by giving the model a way to call external tools. Instead of generating a horsepower number from memory, the model calls a vehicle data tool that returns the verified number from a structured database. The model then uses that response to answer the user. Sourced, not guessed.

How it works in practice

An MCP server is a small program that advertises its tools to the AI agent. For a vehicle data server, the tools might be:

• search_vehicles — find vehicles by year, make, model
• get_vehicle — get full specs for one vehicle
• get_recalls — get recall campaigns for a vehicle
• lookup_dtc — look up a diagnostic trouble code

The agent sees these tools, understands from their descriptions when to use them, and calls them with appropriate parameters. The server handles the API request and returns structured data. The agent incorporates that data into its response.

From the user's perspective, they asked a question and got a sourced answer. From the developer's perspective, they installed an MCP server and the AI gained access to real vehicle data without any prompt engineering or fine-tuning.

The installation problem

Most MCP servers require some setup: install a package, configure environment variables, add a JSON block to your AI client's config. This is straightforward for developers but it's still friction.

CarVector's MCP server installs with one command:

npx -y carvector-mcp --key cv_your_key

Or if your client supports remote MCP servers over HTTP, point it directly at the hosted endpoint with no installation at all:

{
  "mcpServers": {
    "carvector": {
      "url": "https://api.carvector.io/v1/mcp",
      "headers": { "Authorization": "Bearer cv_your_key" }
    }
  }
}

Four tools, one dependency, 150 lines of readable JavaScript, zero telemetry. The code is open source under MIT.