Building AI agents with MCP

Excelrs has two live rows for the same topic: Building AI Agents with MCP. One is the winter internship twin (ZJAHaiLSJAgent, ₹1500, thirty days). The other is the April 2026 training (nz1xJsMCP). Both pages say the same job: learn MCP, then build agents in guided online sessions. Both have ended on the listing. This post teaches the topic those rows name. It is not a day-by-day replay of the sessions.

The problem MCP solves

A chat model is good at language and bad at your life. It does not know the files on your laptop. It cannot add a row to your sheet. It cannot read the PDF you were given in class unless you paste the PDF into the prompt every time.

People work around that by stuffing more text into the prompt. That works for a demo. It fails when the file changes, when the sheet has two hundred rows, or when you want the same agent to work for the next student.

MCP — Model Context Protocol — is a standard way to plug those tools in. The model stays a model. Your filesystem, your browser, your campus API, your notes folder become servers the agent can call. The agent asks for a tool. The server does the work. The model reads the result and continues.

An agent without tools is a talker. An agent with MCP is a worker that can touch your world.

What “agent” means here

On Excelrs pages, an agent is not a personality and it is not a desktop mascot. It is a loop:

1. You give a job in plain language.
2. The model decides whether it needs a tool.
3. A tool runs — list files, read a page, call an API.
4. The model reads the result and either finishes or calls another tool.
5. You see the answer and the trail of what it touched.

That loop is the same idea you meet later in LangGraph weeks and in the AI Agent Lab. MCP is the plug. LangGraph is one way to run the loop. You can learn the plug first.

A picture you can keep

Imagine a student project: “Answer questions only from my internship report PDF.”

Without MCP, you paste pages into ChatGPT. Tomorrow the report has a new section and yesterday’s chat is stale.

With MCP, a small server can search that folder. The agent calls search and read. You can watch those calls. If it cites a heading that is not in the file, you caught a lie.

The protocol sits between the model and your tools. That gap is where most student demos cheat.

What the Excelrs offering covered

The listing is short and we will not pad it. Participants “gain a deep understanding of MCP fundamentals and learn to build AI-powered agents through guided online sessions.” Skills named on the internship twin: AI agents, MCP, prompt engineering, API integration, workflow design. Certificate score 1000 on the winter row. Price ₹1500 on that row. The April training is the same topic on a thirty-day calendar.

That is all the catalog will support. If you need the exact repo from week two, that note is not on the page. Ask for a new batch, or wait until session notes are turned into a lab post.

What you can practise, from the public protocol:

• Run a client that can list MCP servers.
• Connect one boring server — files, or a single HTTP API you already have.
• Give the agent a job that fails without the server (“how many headings are in this PDF?”).
• Read the tool trace. If there is no trace, you are back in chat-land.
• Write down one case where the agent must refuse (delete files, invent a price, skip the PDF).

What good looks like vs a demo

Judges and mentors at Excelrs events look at the second column. So do clients. MCP is interesting because it makes that column possible. It is not interesting as a buzzword on a slide.

How this maps to internships, not just a workshop

The winter row is an internship twin. On Excelrs, internships were folded into trainings; /internships/ZJAHaiLSJAgent redirects here. If you are doing a longer campus internship — diploma Python, diploma AI, a LangGraph summer — MCP can be one module inside that path. It does not replace the campus program and it does not waive those fees.

Gurukul’s public curriculum lists MCP servers next to RAG, tool calling, and workflows. Same family. Different door. This post enrolls you in neither Gurukul nor a new winter batch.

Safety, in plain words

An agent that can read your files can also be asked to send those files somewhere else. An agent that can call an API can be asked to call the wrong one.

Rules we keep on Excelrs work:

• Start with read-only tools.
• Keep the folder small and boring — your own notes, not a client dump.
• Never put passwords, service keys, or other people’s phone numbers in a file the agent can read.
• If a tool can write or pay or delete, it stays off until a person is watching.

That is not a legal lecture. It is how you avoid a bad week.

The refused write is part of the lesson. An agent that can do everything is not a student project.

If the listing says ended

It does. You can still read the offering, share the URL, and learn the protocol from public docs. A new MCP batch will show up as a new date on /trainings. Do not treat this article as an open enroll.

When you are ready for a longer, current lab with a published week plan, the live row is AI Agent Lab. Week 4 there is tools. Week 5 is agents. MCP is one way those weeks get real.

Checklist

• You can say what MCP stands for and what problem it solves.
• You can draw the model, the client, and one server.
• You have one job that fails without a tool.
• You know the Excelrs URLs for the winter twin and the April training.
• You did not invent a session you did not attend.
• You did not move a fee or invent a certificate name.

Build one agent that reads a folder you own. Then stop and write down what it was not allowed to do. That write-up is the start of a real internship project.