요즘 난리난 MCP! 10분안에 정리해줌.

Summary

노마드 코더는 최근 급부상 중인 MCP(Model Context Protocol)에 대해 설명하며, 개발자가 알아야 할 중요한 기술로 강조합니다. MCP는 AI 어플리케이션이 외부 시스템에 연결하는 방식을 표준화한 프로토콜로서, 다양한 AI 도구와 시스템 간의 통합을 용이하게 합니다. 빠른 설치와 사용이 가능하지만 초기 단계이기 때문에 몇 가지 불편함이 있을 수 있습니다. MCP 서버 구축은 새로운 수익 기회를 제공할 수 있으며, 앞으로 더 많은 개선이 기대됩니다.

Highlights

• MCP는 AI 어플리케이션의 외부 시스템 연결을 표준화하여 통합 문제를 해결합니다. 🌉
• Google 트렌드에서 MCP 서버의 검색량이 급증하고 있습니다. 🔍
• MCP를 사용하면 AI 개발자가 같은 구현을 반복할 필요가 없습니다. 🔄
• MCP 서버는 '도구'와 '리소스', '프롬프트'로 구성되어 AI 앱과 상호작용합니다. 🧰
• 미래에는 MCP 서버 연결이 더 간편해질 것이고, 원격 서버 지원과 같은 기능이 추가될 것입니다. 🌐

Key Takeaways

• MCP는 AI 앱 통합의 표준이 되며 개발자에게 중요한 기술입니다. 🚀
• MCP 서버는 다양한 툴과 리소스를 노출하여 AI와의 상호작용을 가능하게 합니다. 🤖
• MCP의 도입으로 UI가 축소되고 AI 중심의 인터페이스가 될 수 있습니다. 🤔
• AI 개발 시장에서 MCP 서버 구축은 새로운 수익 모델을 제공합니다. 💰
• MCP는 여전히 초기 단계이며 향후 많은 개선이 기대됩니다. 🛠️

Overview

MCP, 또는 Model Context Protocol, 은 AI 어플리케이션과 외부 시스템 간의 연결을 표준화하기 위해 Anthropic에 의해 개발되었습니다. 이 프로토콜은 AI 애플리케이션이 시스템 및 데이터 소스와 쉽게 연결할 수 있도록 해, 개발자들이 다양한 구현을 반복하지 않고도 통합할 수 있게 합니다.

MCP 서버는 여러 도구, 리소스, 프롬프트를 통해 AI와 상호작용할 수 있는 기능을 제공합니다. 이러한 MCP의 사용은 AI 중심 인터페이스로 전환을 가속화할 수 있으며, 이는 UI의 필요성을 줄이고 AI가 모든 것을 수행할 수 있는 환경을 조성할 수 있습니다.

현재 MCP는 초기 단계에 있으며, 일부 사용 경험의 불편함이 있습니다. 하지만, 향후 개선을 통해 원격 서버 지원 및 MCP 서버 등록부 등 다양한 기능이 추가될 예정이며, 이는 AI 에이전트가 필요로 하는 MCP 서버를 자동으로 발견하고 연결할 수 있게 만들 것입니다.

Chapters

• 00:00 - 00:30: Introduction to mCP This chapter introduces the concept of model context protocol (mCP), which is gaining popularity based on Google Trends. It emphasizes the importance for developers to learn about mCP and acquire skills to build mCP servers, projecting it as a valuable skill in the near future. The chapter outlines the learning objectives: understanding what mCP is, how it functions, the process of constructing an mCP server, and ways to connect applications like CLA and csor to an mCP server. mCP was developed by Anthropic, known for creating the Cloud, to standardize the connection of AI applications to external systems.
• 00:30 - 01:00: Why mCP is Important This chapter explains the significance of the mCP (Modular Connection Protocol) in AI development. Without mCP, AI application development is fragmented because developers have to create unique implementations for connecting AI applications to different systems and data sources. mCP provides a standard interface, simplifying this process by allowing AI applications to connect seamlessly to other systems without needing to build a new implementation each time. The chapter also describes the basic functionality of mCP, where mCP servers expose data, tools, and prompts, which can be accessed by mCP clients (AI applications).
• 01:00 - 01:30: How mCP Works This chapter explains the operation of mCP, which involves cloud or server connections to the NPC servers. These servers expose tools that are essentially functions running on the server, akin to sending a message or updating a database record, similar to a POST request to an API endpoint.
• 01:30 - 04:00: Building an mCP Server The chapter discusses the process of building a micro-Community Platform (mCP) server. It mentions the availability of over 2,000 community-created mCP servers for various applications such as GitHub, Gmail, Brave search, MySQL server, and Superbase on the Smith.ai platform. The significance of mCP for AI agent builders is highlighted, noting that if mCP technology becomes widely adopted, it will simplify the creation, sharing, and interoperability of tools across different agent frameworks, AI models, and programming languages. From a user perspective, connecting to mCP servers or platforms like GitHub and Superbase can enhance the functionalities of systems like csor.
• 04:00 - 05:30: Connecting to mCP Servers The chapter discusses connecting to mCP servers from a developer perspective, allowing new capabilities such as reviewing pull requests from GitHub or modifying settings from within a chat window. It highlights the potential for developers to build mCP servers that expose account data or tools for users to interact with their servers through AI applications. An example given is creating an mCP server for nomad coders that could let users watch videos and ask questions directly in the chat window.
• 05:30 - 10:30: Future of mCP and Conclusion The chapter discusses the potential of mCP servers as a new revenue stream for indie developers. These servers can perform useful tasks, allowing developers to charge for API keys. On a positive note, this creates new opportunities for entrepreneurs. However, the chapter also explores potential downsides. If mCP becomes prevalent, it might render traditional UIs obsolete, with chat windows becoming the primary interface. This shift could lead AI companies to gather more data on user interactions, potentially creating closed ecosystems similar to Apple's App Store, thereby controlling which mCP servers are accessible.

요즘 난리난 MCP! 10분안에 정리해줌. Transcription

• 00:00 - 00:30 searches for mCP servers on Google Trends are exploding and you my dear developer friend need to know about mCP and how to build mCP servers because I believe it is a skill that will be very valuable in the coming months today you will learn what mCP is how it works how to build your own mCP server as well as how to connect apps like CLA and csor to your mCP server mCP means model context protocol it is created by anthropic the creators of cloud to standardize how AI application connect to external systems
• 00:30 - 01:00 it is like a USBC port for AI apps without mCP AI development is fragmented because everyone is building their own implementation when they need to connect AI apps to other systems and data sources with MPC we have a standard interface that AI apps can use to connect to other systems without having to build a new implementation every time the way it works is simple on one side we have NPC servers that can expose data tools and prompts and on the other side we have NPC clients AI applications like
• 01:00 - 01:30 cloud or csor that connect to the NPC servers and make use of what they expose tools are functions invoked by thei that run on the server like sending a message or updating a database record they are like a post request to an API endpoint resources can be files database records or API responses that the server exposes to give more context to the AI they are like a get request and prompts are templates that will help the user ask better questions to the AI companies like clouds stripe neon and others have already publish their own official mCP
• 01:30 - 02:00 server and if you go to smith.ai you will be able to browse and install more than 2,000 Community created mCP servers for things like GitHub Gmail Brave search MySQL server superbase and more for AI agent Builders like me if mCP takes off it will make it easier to build share and find tools that can be used interchangeably no matter the agent framework AI model or programming language from the user perspective if you connect your csor to the mCP servers or GitHub and superbase for example csor
• 02:00 - 02:30 will now be able to review pull requests from GitHub or modify a setting in your superbase account from within the chat window from the serers developer perspective you can build an mCP server for your users exposing their account data or tools to manage their account to let them interact with your servers from within the AI apps they use I could make an mCP server for nomad coders that lets you watch videos from within the chat window and lets you ask csor questions about the lecture for example mcpl
• 02:30 - 03:00 offers a whole new Revenue stream for Indie hackers you can build an mCP server that does something useful and charge people for an API key to use your servers tools and data from within the chat window that is on the bright side of things but I think there is a dark side to it too in a way if mCP takes off it will make UI Obsolete and the chat window will be the one UI to rule them all that would mean that AI companies will have even more data about you the services you interact and how you use them they could in theory create a world Garden like Apple's App Store and control the mCP servers you can use from
• 03:00 - 03:30 their app that would suck now to build an mCP server you could go and read this spec and implement it all by yourself or you can use one of the existing sdks that you can find in the model context protocol GitHub repo as of today there are sdks for python typescript Java and cotlin we will use the python SDK to build a hypothetical mCP server for stock traders to sell and buy stocks as well as get stock prices from within the cloud and csor chat Windows to get started first we have to install the SDK then creating an mCP server is the
• 03:30 - 04:00 easiest thing in the world we first import the fast mCP class from the mCP server fast mCP module then we create an instance of the fast mCP class with the name of our server and finally we run the server congrats you just created an mCP server of course this server is pretty useless we now need to add some tools resources and prompts to make it useful first let's add a resource by creating a dummy function that will tell us if the market is open or not and we decorate it with the mCP resource decorator remember resources are like
• 04:00 - 04:30 get requests used to get data from the server resources are identified by a URI scheme in this case Market SL SL state if you wanted Dynamic resources you could use variables in the URI scheme notice how the simple variable is defined in the URI and sent as an argument to the function now let's add a couple of tools they are like post requests they should perform an action we create two functions one for buying and one for selling stocks that are decorated with a mCP tool decorator it is important to describe what the tool
• 04:30 - 05:00 does as well as the arguments it receives and the data it returns since this will be used by the AI apps to understand how to interact with the tool finally let's expose a prompt the point of exposing a prompt is to help the user send clearer instructions to the AI model for example if there are some prompts that your company uses a lot long prompts that you worked on for a long time instead of making people copy and paste them every time you can expose them as a prompt let the user select the one they want to use and complete some variables here for example well let's
• 05:00 - 05:30 say that we crafted a prompt that gets the best financial analysis for a stock we create a function that takes some variables and returns The Prompt as a string using the variables we decorate the function with the mCP prompt decorator and we are done now that our server has resources tools and prompts it is time to connect it to clo and csor but before we do that as you can see mCP is just one more example of how rather than replacing developers AI is creating new Industries and career opportunities for us just as there is now a demand for
• 05:30 - 06:00 developers that can build mCP servers the ability to develop with AI is becoming increasingly important in these times of change it is important to find an environment where you can Thrive rather than stagnate now is the time to expand your career into a position that opens you up for more opportunities as a developer it is time to ride the wave of change and shape your future as a competitive developer this couple of months Actually March and April are the busiest months of the year for job openings it is the best time to change jobs which is by wanted the career
• 06:00 - 06:30 platform for developers has created a reward event and a career guide for getting you hired if you apply for a job through wanted and you pass the first round of job applications you will get a nice o on-one reward in neighor points and you will enter a congratulations raffle to win OB Manan wanted has positions at leading AI companies and it also offers free resume coaching for developers click the link below to join the event so as of right now the way we connect to mCP servers is not the best user experience because we are in the very early stages of mCP so things
• 06:30 - 07:00 aren't as polished as they will be in the future if you look at the feature support Matrix you will see the AI apps that support mCP and which features they support the cloth desktop app not the web app supports resources prompts and tools so we will use that and csor supports only tools for now to connect CLA to an mCP server open the CLA desktop app go to the settings click developer and then click edit config that will reveal the location of a file called CLA desktop config.js that will
• 07:00 - 07:30 be empty after you open it in this file is where we should register the mCP servers we want to use I know it isn't sexy or userfriendly and it will change in the future but it is what it is for now here we will write this code the name of the server is stocks mCP server and the command to run the server is UV run main.py on a folder in my document because we are so early for this for now clo desktop does not support remote servers so we need to give Cloud desktop the command to run the server locally in the future we should be able to connect
• 07:30 - 08:00 to a server just by copy pasting the URL UV if you don't know what that is is like an mpm for python I'll make a video about it soon after saving the file we restart the CLA desktop app and we are ready to go when we open Cloud desktop again you will see a connect icon there you can choose to share context with cloth here we can select the market State resource or the analyze stock prompt we also have a dynamic resource but for now cloth desktop does not support Dynamic resources I select the market State resource and ask Claud to tell me the market State clot will use
• 08:00 - 08:30 the resource to answer the question we can click on the resource and sure enough it will have the text returned by the server in a new chat we can now choose to analyze the stock prompt there we are prompted to complete the variables we set the prompt needs I complete the variables and ask clot to analyze the stock he will use the prompt to answer the question we can click on the prompt and see the variables interpolated in the prompt in a new chat we can click on the tools icon where we can see the tools available they aren't selectable unlike resources and prompts that are selected by the user tools are selected by the model I asked the model
• 08:30 - 09:00 to buy 1,000 shares of cloudfare stock it realizes it needs to use the tool it shows me what input it will use on the tool and asks for confirmation after confirmation CLA lets me know the transaction is complete we can see the input as well as the response from the tool we can try to sell shares and it also works to connect to the mCP server using csor is way easier after opening csor settings we go to mCP and then click add new mCP server there we set the name of the server and we can choose if the server should be called from the command or from a URL csor onl Cloud
• 09:00 - 09:30 desktop supports remote servers in our case we select command and we pass the same UV command we used before CLA will run our server and display the available tools buy stock and sell stock now we can open the chat window select agent mode and ask it to buy a th000 shares of clouds slay stock csor will confirm the use of the tool and it will show the input and the response from the tool then we can ask it to sell half the shares and it will work and that's it that is how easy it is to build an mCP server and how unsexy it is to connected
• 09:30 - 10:00 but that will change soon enabling support for remote servers is actually being worked on which will be a game changer since users will be able to connect to a server just by pasting a URL there are also plans to bring oo support So users will be able to authenticate with their own Services instead of using API keys and finally there is a plan to make an official mCP server registry to make Discovery and verification of servers easier once that is completed in theory you could make an AI agent that will automatically discover the mCP servers it needs to work with and connect to them I really
• 10:00 - 10:30 hope mCP takes off and even though they are direct competitors to death I hope open AI Google and DPS all Implement mCP on their app so we can build even cooler stuff see you in the next one bye-bye [Music]