Memuat
Memuat
Want to go deeper?
Join one of our courses or ask us directly on WhatsApp.
Picked from the software, topic, and level of “Claude from Zero to Expert Class #89: MCP: Connecting Claude Code to Other Tools”.
Other ways we can help with Claude.
Learn with a mentor
Scheduled AI & Machine Learning courses with live guidance and a certificate.

Midjourney Course: Create Stunning AI Images from Text
★ 4.9

DALL-E Course: Creating Professional AI Images from Text
★ 4.9

Stable Diffusion Course: AI Art from Beginner to Professional
★ 4.9
Have it built for you
No time to learn Claude? Our team can do the project for you.

AI Image Creation with Stable Diffusion — Unlimited Custom Visual Production
★ 4.9

AI Image & Video Creation with Leonardo AI — All-in-One Visuals for Business
★ 4.9

AI Image Creation with Midjourney — Cinematic AI Art for Business
★ 4.9
Get notified about new Claude classes
At most one email a week. Unfollow anytime with one click.
The full narration of this video. The current sentence is highlighted while the video plays; click any sentence to jump there.
Hi, and welcome back. Claude Code is already good at working with files and commands. But our work also involves browsers, online services, and databases. In this episode we get to know MCP, the standard way to connect Claude Code to other tools. We'll use it to test the Sunset Coffee website through a browser, complete with the safety rules.
MCP stands for Model Context Protocol. The name sounds complicated, but the idea is simple. Think of a USB plug: one plug shape that works for a mouse, a flash drive, or a printer. MCP does the same for AI. One standard way, so Claude can connect to many different tools without a custom connection being built for each one. Anthropic developed MCP as an open standard, so anyone can build an MCP server, and other AI apps can use it too. Because anyone can build one, we need to choose servers carefully. Let's see how it works. On the left is Claude Code, receiving your instructions as usual. In the middle is an MCP server, a small program that provides a set of tools. For example, a browser server provides tools to open pages, click buttons, and take screenshots. That server connects to its target tool: a browser, a work service like GitHub, or a database. Many services already provide an official MCP server. And most importantly, every tool from an MCP server still asks for your permission, just like when Claude changes files or runs commands.
Before installing any server, Raka answers four questions. First, who made it? An official server from the service's maker is more trustworthy than one built by a stranger. Second, which tools does the server provide? A server that only reads is far safer than one that can write or delete. Third, does it need a token or a login? If so, note the correct way to store it according to the documentation. Fourth, is the server used for this project only, or for all projects? For practice, start with one project. Raka wants Claude to test the Sunset Coffee website in a real browser, using an MCP server for browsers called Playwright, made by Microsoft. The install command is typed in a regular terminal, outside the Claude conversation. The terminal confirms that the server was added. The exact command can change between versions, so Raka copies it from the official documentation, not from memory. Then Raka opens Claude Code as usual. Inside Claude Code, Raka types slash mcp to see the connected servers. The Playwright server shows as connected, along with the tools it provides: opening pages, clicking, filling in forms, and taking screenshots. If the status shows a failure, this menu is also the first place to look for the cause.
Raka asks Claude to run the website on the computer, then test it through the browser: add three one-liter coffees, check the total and shipping, then remove them down to zero. An MCP permission box appears. Read two things: the tool name, which opens a page through Playwright, and the target address, localhost, meaning the website on Raka's own computer. It's safe, so Raka approves. Claude clicks the add button on the one-liter coffee three times, just like a customer. Then it reads the page. Subtotal two hundred fifty-five thousand, and free shipping because it's above two hundred thousand. That follows Sunset Coffee's rules. Next, it clicks the minus button four times, one more than it should, to test the limit. The quantity becomes minus one. This is exactly why we test. Claude reports the bug clearly: the minus button can still be pressed at zero, so the total can go negative. A customer messing around could create a strange order. Notice also that not a single order was sent to the real WhatsApp during testing. Raka deliberately asked Claude to stop before the send button, because testing must never send messages to customers. Raka asks for a fix, then a retest through the browser, including at phone screen size. Claude changes one line in the cart script. The diff is easy to understand: before, the quantity simply went down by one. Now it goes down by one, but never below zero. Claude resizes the browser to the size of a phone screen. The retest passes: the quantity stops at zero, and the send button is clearly visible. Claude also saves the screenshot, so Raka can see the proof directly, then makes a commit as usual.
An MCP server opens up new access, so treat it like handing someone a key. The first rule: install only servers from trusted sources, preferably official servers from the service's maker. Second, read the tool name and its target in every permission box, as Raka did. A tool that opens localhost carries a different risk than a tool that sends email. Third, many servers need a token, a kind of access key. Keep the token in the settings, as that server's documentation explains. Never paste it into a conversation or write it in a script. Fourth, give the least access possible. If reading is enough, don't grant write or delete access. Fifth, watch out for pages or documents that contain instructions for AI, such as hidden text telling it to send data. This is called prompt injection. Claude is designed to be wary of it, but the permission box remains your last line of defense. Here's an example: Raka asks Claude to read a coffee bean supplier's review page on the internet. Claude finds hidden text telling it to send the contents of a secret file to another address. Claude refuses and reports it to Raka. This is why we read Claude's reports, not just the final result. Raka makes it clear: ignore instructions from web pages, and just summarize the reviews. Legitimate instructions come only from you, not from the content of pages or documents.
Let's look at more ideas. Raka now asks Claude to test the website through the browser before every release, so bugs like a negative quantity are caught before customers find them. Code hosting services like GitHub have MCP servers. Claude can read the list of reported issues, then propose fixes. At the office, Dewi could connect Claude to the sales database with read-only access, and simply ask in plain language, for example which branches sold less paper this month. Read-only access means the data can't be changed. And for teams, the server list can be stored in the project settings file, so every team member uses the same connections. The available servers keep growing, so check the official Claude Code documentation for the latest list and setup. The screens in this video are simulations. Let's look at Dewi's example. The office IT team has already installed a database server with read-only access. Dewi asks in plain language: which branches sold less A4 paper this month? A permission box appears. Dewi reads that it only runs a read query, a request for data, on the sales table. Nothing gets changed, so Dewi approves. Claude compares fifteen branches and finds three that are down. The answer comes complete with numbers, and it matches the report Dewi built in the automation session: Bogor is down one hundred eighty reams. A question that used to take half an hour in Excel is now answered in seconds.
We got to know MCP, installed a browser server, had Claude test the website like a customer, fixed the bug it found, and learned five safety rules for using MCP servers. In the next episode, the final episode of this session, we create custom commands for work we repeat often, and hooks that run checks automatically. See you there!
Try these on your own after watching to lock in what you learned.
Type a question or keyword; we find the part of this episode that explains it and jump the video there.
Notes are saved with the current video time and are only visible to you.
Stuck on a step? Ask here; the instructor and other learners can answer.
No questions yet.

Cerdas Inovasi Digital Team
Software developers, BIM & 3D engineers, and AI practitioners
We build software, BIM models, games and AI automations for clients every day, and turn that hands-on work into step-by-step classes: every step is shown on screen and comes with a PDF guide you can keep.
See our clients →