AI agents and MCP: how to give an agent tools

The short version: an AI agent is a model that does not simply answer but carries out steps by itself — searches, calls tools, checks the result and decides what to do next. MCP is the standard way to give an agent those tools. The application that connects to MCP servers and runs that loop is called an MCP client. Below: how it is put together, and three ways to build an agent of your own.

Chat, agent and MCP client — what the difference is

The thing worth remembering: what makes a model an agent is not MCP but the loop around it. MCP is only responsible for that loop having hands — and for not having to sew them back on for every new application. More on the protocol itself is in the overview of MCP servers, and more on agents in general, with no MCP in sight, in the guide to AI agents.

How an agent works with tools

1. The agent receives a task and sees a list of tools with their descriptions.
2. It decides which tool is needed first and calls it — for example web-search.
3. It reads the result and decides again: if the data is enough, it answers; if not, it calls web-crawl on the best link, or another tool.
4. It repeats until the task is solved and returns the outcome: text, a table, a picture, a file.

Route 1. An agent in Claude, Cursor or Codex

The quickest path: these clients already are agents. Connect MCP servers and Claude starts searching the web by itself, Cursor reads the documentation before editing code, and Codex checks against the changelog. Step by step: Claude, Cursor, Codex, OpenCode.

This suits you when the agent is for you personally and interactive: you set a task, watch, correct.

Route 2. An agent in n8n

If the agent has to work on its own — on a schedule, on an incoming email, on a new CRM lead — it gets built inside an automation system. In n8n there is an AI Agent node for that, with a model and tools attached to it. An MCP server is added as a tool through the MCP Client Tool node: you give it the server address and the sign-in method — for key-based servers, the Authorization: Bearer … header.

The reverse also works: the MCP Server Trigger node turns an n8n workflow into an MCP server, which can then be connected to Claude or Cursor. That is how your internal processes — an export from CRM, sending an invoice — become an agent's tools.

Route 3. A ready agent chat, no config files

If you want an agent without JSON and a terminal, you can assemble one in Twelver. The same 26 sets of tools that are available as MCP servers are enabled here with checkboxes in the agent editor — search, image and video generation, voice-over, documents, presentations. And there is something here a plain MCP client does not have:

• Memory — the agent remembers facts between chats and keeps its own archive of material.
• Self-scheduling — the agent runs itself later, once or on a schedule, and sends you the result.
• Knowledge — documents attached to the agent, which it answers from.
• Teams of agents — a lead agent hands tasks to subagents; available on the Premium plan.

Tick the tools below and press "Try it live" — the agent editor opens with that set already chosen.

{
  "mcpServers": {
    "twelver-image-generation": {
      "url": "https://twelver.ru/api/mcp/v1/image-generation"
    },
    "twelver-web-search": {
      "url": "https://twelver.ru/api/mcp/v1/web-search"
    },
    "twelver-document": {
      "url": "https://twelver.ru/api/mcp/v1/document"
    }
  }
}

Which route to choose