IntegrAuth Academy
Identity, security and practical AI — free, from zero. Byte-sized lessons, real diagrams, hands-on labs, zero fluff. Two programs, each with its own exam and certificate.
English is hard? Your browser can translate this page into Tamil, Telugu, Kannada, Malayalam or Hindi. Tech words stay in English so you learn the real names.
How?
- Chrome on phone: ⋮ menu → Translate
- Chrome on computer: right-click → Translate to …
- Edge: the translate icon in the address bar
MCP: plug tools into AI
Priya wants her coding assistant to read team notes and list open issues. She adds two lines to a settings file. Your goal: plug tools into an AI app safely.
🧪 Interactive lab — enable JavaScript to play with this one.
- A system describes its tools once, as a server. Any MCP app can use them.
- The model never talks to a server. The host sends every call.
- Give each server only what it needs. Keep tokens out of shared files.
One plug for every app
Many AI apps want to use many systems, like files and calendars. Connecting each pair by hand is too much work. With MCP, the Model Context Protocol, a system describes its tools once, as a server. Then any AI app that speaks MCP can use them. It is like USB: any USB mouse works in any laptop.
Host, client and server
| Role | What it is |
|---|---|
| Host | The AI app you use, like Claude Desktop, Cursor or ChatGPT |
| Client | A connector inside the host, one for each server |
| Server | A program that offers tools, often on your own laptop |
The model only asks. The host asks you for permission. Then it sends the call through that server's client.
Tools, resources and prompts
| Kind | What it is | Who picks it |
|---|---|---|
| Tools | Actions, like create_issue | The model. The host may ask you first. |
| Resources | Read-only data, like a file | The app |
| Prompts | Ready-made instructions, like /summarize-pr | You |
Now follow one tool call.
🧪 Interactive lab — enable JavaScript to play with this one.
Go deeper: local or remote
A local server is a program on your computer. The host starts it. They pass messages through the program's text input and output. This is called stdio. It is good for files and local code.
A local server runs with your permissions. Whatever you can read or delete, it can too.
A remote server is a web service. A company runs it, and many users share it. The host reaches it over the web. This is called Streamable HTTP. You usually sign in with OAuth, or you give it a token.
Clients and servers send small JSON messages, in a format called JSON-RPC. tools/list asks a server what it can do. tools/call runs one tool.
Go deeper: set one up
Most hosts read a small JSON settings file. The top key is different in each host. Claude Desktop, Claude Code and Cursor use "mcpServers". VS Code's own file, .vscode/mcp.json, uses "servers". Claude Code reads .mcp.json in a project.
Each server gets a name. A local server gets a command to start it. A remote server gets a URL. Here is a local file server that can only see one folder:
"notes": {"command": "npx", "args": ["-y", "@modelcontextprotocol/server-filesystem", "/home/priya/notes"]}
Here is a remote one: "repo": {"type": "http", "url": "https://mcp.example.com/mcp"}. Then reload the host. It lists the tools it found.
- Keep secrets out of shared files. A project's
.mcp.jsonoften goes into the code repository, the shared project folder. Put tokens in environment variables. These are settings kept outside the file. Then point to them, like"Bearer ${GITHUB_TOKEN}", if the host supports it. - Keep it small. Point a file server at one folder, not your home folder. Use a read-only token if you only need to read.
- Approve on purpose. Most hosts ask before they trust a new server. Many also ask before they run a tool. Read those questions, and check your settings.
Priya's first try pointed the file server at /, the whole computer. Her token sat in the shared file. A reviewer caught both. Now the server reads only ~/notes, with a read-only token.
Go deeper: what MCP is not
- Not a model or an agent. It only connects an AI app to tools. See Agents, subagents and skills.
- Not a safety promise. A standard plug says nothing about what is plugged in. A local server is someone's code on your computer. A tool description can hide orders for the model. See prompt injection.
- Not the whole security story. Each system still checks who may do what. See Governing MCP, in the other program.
- Not always needed. For one app with three tools, plain tool calling is simpler. MCP helps when tools should work in many apps.
- Not agent-to-agent. Agents that hand work to other agents use other standards, like A2A. See Agent-to-agent delegation.
Go deeper: history
Anthropic first published MCP in November 2024. In December 2025, it moved to the Agentic AI Foundation, a neutral home under the Linux Foundation. Most major AI apps support it, some more fully than others.
Real servers, as of October 2026: the reference servers, like filesystem, git and fetch, are learning examples. Others include GitHub's official server and Microsoft's Playwright server for browsers.
Adding a server means installing software. A server from a random post runs a stranger's code with your permissions. Use official or well-known servers, and only the ones you need.
Run the free MCP Inspector. It needs Node.js. Connect it to the reference filesystem server, pointed at an empty folder. List the tools and call one.