MCP for Beginners: Connect Your First Tool in 10 Minutes

The quickest way to understand the Model Context Protocol is to connect one server and watch your assistant use it. You need three things: an assistant that supports MCP (Claude Code, Claude on the web, Cursor, VS Code with Copilot and others do), the address of a remote MCP server, and a browser to sign in with. The whole job is: add the server, approve the sign-in, ask the assistant who it is, give it one small task, then check the record of what it did. This guide walks through those steps with a task board as the example tool, because every step is easy to see.

If you want the concept first, read what MCP is or the one-paragraph glossary entry. This page assumes you would rather learn by doing.

The only three words you need

MCP’s own architecture overview↗ (opens in a new tab) names three participants. You will meet all three in the next ten minutes, so it helps to know which is which.

• The host is the AI application you type into: Claude Code, Cursor, VS Code, Claude on the web.
• The client is the connection the host opens for each server. You never touch it; the host creates one per server you add.
• The server is the program that offers tools. A local server runs on your machine over the stdio transport; a remote server runs on the internet over Streamable HTTP. For a first try, remote is simpler: there is nothing to install.

A server offers tools, and each tool has a name, a description and the inputs it takes. When you ask the assistant for something, the host shows it the list of tools and the model decides which to call. That is the whole trick. The assistant is not “logged in to your app”; it can call the specific tools the server offers, with the permissions you granted.

Minute 0 to 2: pick one server and find its address

Choose a server that does something you can check with your own eyes. A task board is a good first choice: when the assistant adds a task, you can open the board and see it. The fenbs server is one address, and it is the one used below.

https://fenbs.ai/api/mcp

Before you connect anything, sign in to the tool itself in a browser and add one task by hand. An empty board makes it hard to tell “the connection failed” from “there is nothing there”.

Minute 2 to 5: add the server to your assistant

Every host has its own place for this, but the information you give it is the same: a name for the server and its URL. Pick the line for the assistant you use.

claude mcp add --transport http fenbs https://fenbs.ai/api/mcp

In Claude Code, the server is saved for the current project unless you add --scope user (every project on your machine) or --scope project (a shared .mcp.json file in the repository), as the Claude Code MCP documentation↗ (opens in a new tab) explains. Then run /mcp inside Claude Code and choose the server to sign in.

{
  "servers": {
    "fenbs": {
      "type": "http",
      "url": "https://fenbs.ai/api/mcp"
    }
  }
}

In VS Code you can create that file by hand or run “MCP: Add Server” from the Command Palette; the VS Code MCP servers page↗ (opens in a new tab) covers both, and VS Code asks you to confirm you trust a server before it starts it.

• Cursor: add the same URL under mcpServers in mcp.json, either .cursor/mcp.json in the project or the one in your home folder. The Cursor page has the snippet.
• Claude on the web or desktop: open the connectors settings, add a custom connector and paste the URL. Custom connectors depend on your plan; see Connect Claude.
• ChatGPT, Devin Desktop (formerly Windsurf), Gemini CLI, Codex CLI: the same URL goes in each tool’s MCP or connector settings. Each has its own page under integrations.

Minute 5 to 6: sign in and approve

The first time the host talks to a remote server that needs a sign-in, the server answers “not authorised”, and the host opens your browser. This is standard behaviour: the MCP authorization specification↗ (opens in a new tab) builds on OAuth 2.1, the same kind of sign-in you use when an app asks to “continue with” another service. You sign in on the tool’s own page, never inside the assistant, so the assistant never sees your password.

On fenbs, the approval page shows the name the app gave itself (marked unverified, because anyone can call themselves “Claude”), where you will be sent back to, and the board it will work in. Reading the board is always on; “Add and change tasks” and “Comment” are yours to tick. For a first try, tick all three so you can see a write happen, then press Approve. The assistant receives a token it keeps; you never copy one.

Minute 6 to 9: three first prompts

These three are chosen to test one thing each: that the connection works, that reading works, and that writing works. Type them as written, one at a time, and watch the tool calls the host shows you between the prompt and the answer.

1. “Call fenbs_whoami and tell me who you are acting as, which boards you can see, and your role on each.” This proves the sign-in worked and settles what the assistant is allowed to do before it does anything. If it answers with your name and a role, you are connected.
2. “List the tasks in To Do and Next Up, with their refs and kinds, and tell me which one looks oldest.” The assistant should call fenbs_list_items and answer with refs such as FET-003 or BUG-007. Compare its list with the board in your browser; they should match exactly.
3. “Add a task called ‘Tried MCP for the first time’ to To Do, kind enhancement, then comment on it saying which tools you used.” That is fenbs_create_item and fenbs_comment. Most hosts pause and ask before a tool changes something; read what it is about to send, then allow it.

If the third prompt is refused, that is useful too. A fenbs refusal names the permission that was missing and the role the assistant holds, so the assistant can tell you in a sentence why it stopped. The usual cause on a first try is leaving “Add and change tasks” unticked.

Minute 9 to 10: check what it did

Trust comes from checking, not from the assistant’s summary. Open the board in your browser, then its History page. Every change is listed with who made it, and changes made through an assistant carry its name, for example “Claude via Sam Ade”. Choose AI Assistants in the filter to see only those lines. You should find exactly one new task and one comment. If you see more than you asked for, you have learned something important about that assistant before it mattered.

Then find the switch. The token the sign-in created is listed by name under Settings, “Connect an AI assistant”. Revoking it stops the assistant at once and leaves your own sign-in alone. Knowing where that is before you need it is the last step of a first connection, not an afterthought.

If something does not work

• The assistant does not mention the tools at all: check the host lists the server as connected (/mcp in Claude Code, the MCP view in VS Code). Some hosts need a new chat after you add a server.
• The browser never opened: start the sign-in yourself from the host’s MCP list. In Claude Code that is /mcp, then the server.
• The assistant answers from memory instead of calling a tool: name the tool in the prompt, as prompt 1 does. Once it has used a server once, it usually reaches for it without being told.
• You are building your own server and want to test it without an assistant: the official MCP Inspector↗ (opens in a new tab) lists a server’s tools and lets you call them by hand. For a server you only use, debugging MCP tools covers the assistant side.

What to do in the next ten minutes

Once one server works, the pattern repeats for every other: name, URL, sign-in, a “who am I” prompt, one read, one write, a check. Add servers one at a time, and give each only the permissions the job needs. If you do not write code, MCP without coding goes further on the settings screens in Claude and ChatGPT. If you use Claude Code every day, connecting Claude to your project board covers the full tool list and what a refusal looks like.