A deliberately non-technical session. No code — just proof that this works, and the vocabulary you’ll need when we start building next week.
1. The Demo: An Agent That Actually Booked Something
The goal was a real action, not a description of one: book a table for four in Hyderabad.
- Open Claude Desktop and add a custom connector.
- Point it at the Swiggy Dineout MCP endpoint:
https://mcp.swiggy.com/dineout - Sign in with a phone number and OTP to link the Swiggy account.
- Ask in plain language: a table for four, Financial District, this Saturday at 8 p.m.
- The agent searched, offered options, and confirmed a booking at Red Rhino.
No app, no form, no browsing. One sentence.
Before you try it: connector availability varies by plan and region, so check what your account supports rather than assuming.
2. What MCP Actually Is
MCP is an open standard that connects AI applications to the systems, data and actions they need.
The analogy that sticks: a USB-C charger for AI. Build your tool once as an MCP server and it works with any MCP-capable client — Claude, ChatGPT, others — with no rebuild per vendor.
The one condition: the client has to support MCP. The standard spreads as clients adopt it, not automatically.
The Real Architecture
Learn these three words, in this order:
| Part | What it is |
|---|---|
| Host | The AI application the user is in — Claude Desktop, an IDE, your own app |
| Client | Lives inside the host; one client per server connection |
| Server | Your tool. It exposes capabilities and does the work |
(The “Model, Context, Protocol” expansion explains the name, not the design. The design is host, client, server.)
What a Server Exposes
- Tools — actions the model can take (book a table, create an issue)
- Resources — data the model can read
- Prompts — reusable templates
Local or Remote
- Remote servers talk over HTTP, which is how the Swiggy connector works, and why it needed a proper sign-in.
- Local servers run on your own machine over stdio. That’s what you’ll build first in Python.
3. The Agent Decides the Order
You describe the goal; the model decides which tools to call and in what sequence. You can influence it through prompting, but you don’t script it. That’s the difference between an agent and a workflow.
And the standing rule still holds: the model names the tool, something else runs it.
4. The Ecosystem
MCP servers already exist for Slack, GitHub, AWS, Azure and Google Cloud. The Playwright server is worth a look — it drives a real browser, which makes UI testing and web automation available to an agent.
5. Why a Business Should Care
This is the strategic part.
If your e-commerce or registration system has an MCP server, customers can reach it from whichever AI assistant they already use. You stop demanding that people come to your app, and your service becomes available wherever they’re already asking questions.
That’s a distribution argument, not a technology one — and it’s why adoption is moving quickly.
6. Where We’re Going
This session was the what and the why. From next week we stop being users and start being builders: the architecture of an MCP server, and the code to create one.
The target: custom MCP servers that let agents work with your own organisation’s applications.
✅ Action Items
- Joined late? Review the previous three to four foundation sessions — the summaries are all on this site.
- Try an existing MCP connector in a client you already have, and watch which tools it calls.
- Think of one system at work you’d want an agent to reach. That’s your candidate for a first server.
⏭️ Next up: MCP server architecture, and your first server in Python.
1 comment