MCP: Understand It, Set It Up, Use It

0:00AI has gotten really good but you are

0:02often still the blocker for software

0:04development specifically you copy paste

0:06things code SQL queries deployment

0:09errors and then the AI suggested fixes

0:12and actions run this deploy that get me

0:15the logs so the iteration loop is slow

0:17because you are the bottleneck and there

0:19is no real added value here you're

0:22mostly doing monkey copy paste work the

0:25AI is smart enough to run these things

0:27itself do a deployment check the logs,

0:30create the database, iterate until it

0:32works, and that's what MCP [music]

0:34enables. In this video, we'll understand

0:36what MCP is and go through a quick

0:39hands-on setup. You'll see that the

0:41magic happens when you chain multiple

0:43MCPs with the knowledge of an LM. So,

0:46MCP stand for model context protocol.

0:49It's a standards that lets AI tools

0:51connect to external services. Say you

0:53want to search and summarize all notion

0:56pages related to a topic. Typically you

0:59would search yourselves then copy paste

1:01it and say give me a one-pager summary.

1:04With the notion MCP the AI can read your

1:07documents directly and even creates new

1:10ones. No downloading no copy paste. You

1:12just say summarize all your work that

1:14has been done on a topic example DAX and

1:17it does it. Same with databases. The AI

1:19writes you a query you run it. It fails.

1:22You copy the error back get a fix. It's

1:25slow. With an MCP, the AI can run the

1:27query itself, see it fail, fix it, and

1:29keep iterating until it works. And it's

1:31not just about avoiding copy paste. It's

1:34about the AI can now act on tools

1:36directly, see what's happening, and keep

1:39trying until it succeed. So that

1:41feedback loop is the superpower. MCP was

1:44created by Entropic, the company behind

1:46Clo in November 2024. The idea was

1:49simple. Every AI tool was building its

1:52own notion integration for example its

1:54own Slack integration it's on everything

1:56chat GPD as one cloud as one. So MCP

1:59said let's make a standard the community

2:01or mostly the owner here notion for the

2:04notion MCP example builds one MCP server

2:08and all cloud chat GPT cursor whatever

2:11can directly connect and use it as a

2:14superpower tool. Something important to

2:16mention in December 2025 Entropic

2:18donated MCP to the Linux Foundation

2:21specifically the new Agentic AI

2:23foundation the same foundation that

2:25stewards Kubernetes and PyTorch OpenAI

2:28Google Microsoft AWS all joined as

2:31founding members. So this is now

2:32becoming slowly the industry standard

2:35for LMS to connect through tools. But

2:38what exactly is the protocol definition

2:40and you want to stay with me because as

2:43a user you'll give access to various

2:45things so better understand which power

2:47you give to your AI before they take all

2:50over. So when you install an MCP server

2:52you're giving your AI access to specific

2:55capabilities like instead of just

2:57writing back text it can perform various

2:59action. This can come in three flavors

3:02and you'll see them when youize the

3:04connection. First tools. These are

3:06action the AI can take. For notion MCP

3:09server, it has tools like notion fetch,

3:12notion get comments. These are readonly

3:14tools. But of course, you have also

3:16another category here, write and delete.

3:19When the AI needs to do something, it

3:21calls a tool and you can specify which

3:24permission per default it has. Always

3:26allow, need approval or block. The

3:28second is resources. This is data the AI

3:31can read. A file system MCP expose your

3:34files as resource. A database MCP might

3:37expose your schema as a resource. A CRM

3:40like HubSpots might expose your contact

3:42list. Resources are readonly context and

3:45the AI can see them but can't modify to

3:48resource. That's what tools are for. And

3:51the third one finally is prompts. Think

3:53of these are slash commands or shortcuts

3:55and need boosts to your request that

3:58provide better context. A database MCP

4:00might offer a analyze table prompt that

4:03automatically structures how the AI

4:06should analyze your data. For instance,

4:08first getting the schema, etc. Honestly,

4:11most MCP servers focus on tools. Prompts

4:14are nice to have, but not essential. But

4:16now you know what you're looking at when

4:18you authorize a connection. Note also

4:20that MCP released just a new

4:22specification called MCP apps, but they

4:25are entirely something else. Tell us in

4:27the comments if you want a dedicated

4:29video specifically on that. Now here is

4:32where people get confused. There are two

4:34type of MCP servers remote and local.

4:38Remote servers run in the cloud. Modduck

4:40runs theirs. Notion runs theirs. Linear

4:43Slack. They all their own MCP servers.

4:46You connect via HTTPS. Authenticate with

4:50OS. Click sign in with Google styles and

4:53you're done. No installation, no config

4:55files. in Clo these are called

4:57connectors and in chat GPD they are

5:00called apps now. Yeah, don't ask me why

5:02they couldn't just call it MCP. Maybe

5:04for folks that have no idea what MCP is,

5:07but you're not one of those. Not

5:09anymore. Anyway, local servers run on

5:12your machine and when you configure one,

5:14you're basically starting a small web

5:16server locally that translate the AI

5:19request into action on your system. The

5:22file system MCP for example runs on your

5:25computer and lets the AI read and write

5:27files in your folder specifically. So if

5:30you're using a local model through

5:32Olama, the data never leaves your

5:34laptop. One important security tip,

5:37don't install MCP servers that aren't

5:40open sourced or backed by the main

5:42company of the product. A local MCP runs

5:45codes on your machine with your

5:47permission. And if it's a random GitHub

5:49repo with no star and no company behind

5:52it, skip it. Stick to the official

5:54servers or wellknown open-source

5:57projects that you can actually inspect.

6:00Stay safe. All right. So, how to set

6:02them up. Okay. So, no more copy paste

6:04monkey works. Two ways to add MCP

6:06server. Let's show you both. One way is

6:08to go and click the like button and

6:10subscribe and ta your MCP is installed.

6:13I'm just kidding.

6:14 >> Or no. So for remote servers, the

6:16easiest way is through the approved

6:18directory. These are servers that have

6:21been reviewed and trusted by the AI

6:23platform. Would it be cloud or chat GPD?

6:25In cloud, you go to settings,

6:27connectors, browse the directory. You

6:29can find moduck, click connect,

6:30authenticates, done. You're granting the

6:33AI permissions to use that service on

6:35your behalf. Same for notion. Browse,

6:38connect, authorize. Now I have both

6:40moddion

6:42connected. The second way to install an

6:43MCP server is through a JSON

6:46configuration. This works for both

6:48remote servers that aren't approved in

6:50the directory and local servers. For

6:52remote servers, you'll see that in the

6:54configuration, you have to connect to an

6:56endpoint that is hosted by the owner.

6:59For local servers, you're running the

7:01servers directly. So, it would be

7:02classically an npm command or a UV

7:05command to bootstrap the local servers.

7:08So that means you either need NodeJS

7:10with npm for JavaScript servers or

7:13Python with UV for Python servers and

7:15most MCP servers are built in one of

7:18these two. You can check mcp.so over

7:2117,000 servers are there. All right,

7:24let's see what it actually enables. I

7:26worked in Devril at MoDoc and part of my

7:29job is tracking what people say about us

7:31and moduck on various things like a news

7:34you know helping people when there is

7:36confusion. So normally you search your I

7:38can use open threads read the comments

7:41take notes copy do but you actually can

7:43do that in just one prompt. I've

7:45connected two MCP servers motoduck which

7:48has a public data set with the entire

7:50acronuse history more than 50 millions

7:53post and notion for creating documents.

7:55So watch what happens. So this is my

7:57first initial prompt not that much

7:59complicated asking to create a report of

8:02comments that I can act on. So first

8:04it's going to call moduck running SQL uh

8:07across the 50 million rows data set. It

8:10finds the post mentioning DDB and moduck

8:13get the top discussion pulls the

8:15comments. Now here is where the LLM does

8:17its thing. It's not just moving the

8:20data. It's reading these comments and

8:22understanding them. Oh this question is

8:24about S3 caching and it's not answered.

8:26This one is a misconception about scale.

8:28We should correct it. This is a filter

8:30request about DAX. And now it calls

8:33notion to create the report. One prompt,

8:36two MCP servers. So now you see the

8:38power of combining two MCP with an LLM.

8:41And I have a document ready to share

8:43with my team. That's it. Now go play

8:45around. Install one MCP, the Moduck one,

8:48the notion or whatever fits you. See

8:51what you're doing as a copy paste and if

8:53there is actually an existing MCP. And

8:56yeah, now get the hell out of here and

8:58go build something. Cheers.