[0:00]Hi guys, my name is Nitish and you're welcome to my YouTube channel. So, agar aapne mera last video dekha hoga, toh aapko yaad hoga ki maine is channel pe ek naya playlist start kiya hai, aur is playlist ka naam hai Model Context Protocol, MCP. Aur is playlist ko start karne ke piche jo mera motivation hai, woh yeh hai ki yeh jo particular term hai, MCP, yeh pichle ek saal mein bahut zyada popular hua hai. Aur aisa lag raha hai ki agle teen se paanch saalon mein MCP will become an industry standard. Which basically means ki industry mein har kisi ko MCP apne softwares mein integrate karna padega. And that is why mujhe aap logo ki side se bahut zyada messages aur response aaya hai ki sir aap please MCP cover karo. So, that is why pichle 30-40 dino se I was studying MCP, I was trying to implement it on my laptop aur eventually maine kya kiya ek curriculum plan kiya. Ab us curriculum ke hisab se I have decided ki main is playlist mein teen videos dalunga. Jisme se pehla video hoga The Why, dusra hoga The What, tisra hoga The How. Aur mera goal yeh rahega ki in teen videos ke through main aapko in-depth tarike se MCP sikhaun. Ab agar hum baat kare aaj ke video ki, toh aaj ka video is playlist ka pehla video hai. Aur isme hum cover karenge The Why aspect of MCP. So basically, mera goal yeh hai is video mein ki main aapko in-depth tarike se sikhana chahta hoon ki MCP ki zarurat kyon padi. Aisi kaun si problem hai jo MCP solve karke deta hai. Aur yeh poori cheez ko samjhane ke liye main ek bahut story-telling approach lene wala hoon. Main ekdum shuru se shuru karunga jab ChatGPT picture mein aaya aur wahan se aaj ki date tak jo bhi cheezein hui hain, main woh sari cheezein aapko ek story-like fashion mein bataunga. Aur isse yeh fayda hoga ki main aapke dimag mein bahut deeply ek intuition develop karwa paunga ki MCP ki as a technology zarurat kyon hai. theek hai? Toh, I really hope main aapko pehle ek minute mein samjha paya ki is video mein hamara goal kya hai. Aur ab agar aap ko yeh sab kuch sahi laga, sab kuch sahi sound kar raha hai, toh now let's start the video. So guys, humari kahani ki shuruat hoti hai 30 November 2022 ko, jis din ChatGPT release hua tha. Aur within five days, ChatGPT ne 1 million user ka mark cross kiya. Aur agle do mahino mein 100 million user ka mark cross kiya. Aur yeh koi chote-mote numbers nahi hain. In fact, ChatGPT ke pehle kisi bhi software ne itne crazy numbers achieve nahi kiye the. Bhale hi aap baat kar lo Google ki, ya phir Facebook ki, ya phir Twitter ki. In fact, ChatGPT is a completely different class of software. Aisa software jo pehle kabhi aaya hi nahi. Aur ChatGPT aapko ek aisi capability deta hai jo pehle kisi software ne aapko nahi di hogi. Aur woh capability hai ki aap mashino ke saath bilkul waise hi baat kar sakte ho jaise aap kisi insaan ke saath karte ho. In your natural language. Soch ke dekho, humara mashino ke saath jo relationship hai, woh shayad 500-600 saal purana hai. Pehle mechanical machinein hua karti thi, phir electrical machinein aayi aur phir pichle 50 saalon mein hum computers ke saath bahut kaam kar rahe hain.
[3:27]But in 500-600 saalon mein humara jo relationship raha hai woh mostly transactional raha hai. Which basically means ki agar mujhe apne machine se kuch bhi chahiye, toh main simply ek action perform karta hoon aur mujhe uska palat ke result mil jata hai. Mujhe garmi lag rahi hai, mujhe fan on karna hai, main jaa ke switch daba dunga. Mujhe kuch calculation karna hai, main calculator pe buttons daba dunga. Mujhe ek form fill karna hai, main simply keyboard pe cheezein type karke ek button press kar dunga. Aur machine humara kaam kar degi. This is a transactional relationship. Jahan pe hum bahut minimal tarike se baat karke apna kaam nikalte hain. But ChatGPT ke aane ke baad aap computers ke saath bilkul waise baat kar sakte ho jaise aap kisi insaan ke saath karte ho. Aap chaho toh apne aap ko express kar sakte ho, saamne se machine bhi khud ko express kar sakti hai. Aap apne computer ke saath kisi thoughtful discussion mein engage kar sakte ho. Aap in fact usko apna work partner bana sakte ho. Yeh sab kuch possible hua ChatGPT ke aane ke baad. That is why I am saying ki ChatGPT is a completely different class of software. theek hai? Ab kisi bhi aur software ki tarah, ChatGPT jab launch hua toh uska adoption dheere-dheere hua. In fact agar aap mujhse poocho, toh ChatGPT ka jo adoption hai woh teen stages mein hua. So, jo pehla stage hai, usko main bula raha hoon The Stage of Pure Wonder. Mujhe abhi tak yaad hai, uh around December first week mein mere ek student ne mujhe WhatsApp pe message kiya aur usne bola ki sir, aap ek baar yeh particular chatbot try out karo. It's crazy good. Aur mujhe abhi tak yaad hai ki agla poora din main ChatGPT ke saath baat kar raha tha. And I was shocked ki yeh chatbot kitna intelligent hai.
[5:20]And I'm pretty sure aap sab ne yeh same cheez experience ki hogi jab aapne ChatGPT ke saath first time baat ki hogi. Toh is poore ke poore stage mein mujhe lagta hai logo ne sirf ek kaam kiya. Aur woh tha apne andar ki curiosity ko shant kiya. So basically, logo ne ChatGPT se ut-pataang sawal poohe, jaise ki I am pretty sure kisi ne poocha hoga ki Quantum Physics ko explain karo kisi cat ke perspective se. Ya phir kisi aur ne poocha hoga ki agar gravity ulti ho jaye toh kya hoga. Ya phir kisi ne poocha hoga ki pizza ke upar ek gana likh ke batao, Shakespeare ke style mein. Aur basically hum yeh crazy questions pooch rahe the aur ChatGPT intelligently respond kar raha tha aur hum uska screenshot leke LinkedIn, Instagram, WhatsApp pe post kar rahe the. So in short in this first wave of adoption, hua kya ki social media poora exploded. Aur hume constantly yeh screenshots dekhne ko mil rahe the ki logo ne kya question poocha aur ChatGPT ne kya reply kiya. So yahan par kuch meaningful kaam nahi hua, but humari jo curiosity hai woh thodi shant hui. Phir aata hai second wave of adoption. Isko main professional adoption bula raha hoon because ek baar jab woh initial period khatam hua, humari curiosity thodi shant hui aur hum thoda samajh gaye ChatGPT ko, toh humare andar ek question automatically aaya ki yeh jo masti hum kar rahe hain ChatGPT ke saath yeh toh theek hai. But kya yeh chatbot itna intelligent hai ki hume humare professional kaam mein help kar paye. So this was the first time jab kisi lawyer ne jaa karke ek 50-page ka contract dala ChatGPT mein aur bola ki yaar yeh summarize karke do. Aur ChatGPT ne kar dikhaya. Maybe this was the first time jab kisi coder ne apna code paste kiya aur bola ki yahan pe error aa raha hai, Can you debug this? Aur ChatGPT ne karke dikhaya. Ya phir mere jaisa koi teacher gaya aur usne bola ki yaar mujhe curriculum plan karna hai aur mujhe sab kuch padhne ka mann nahi kar raha hai, Can you do this for me? Aur ChatGPT ne phir se karke dikhaya. So this was the first time jab hum sabne collectively realize kiya ki yeh sirf hasi-mazaak karne ka tool nahi hai. In fact, this chatbot has got serious potential and it can become our work partners. Matlab, hum chahein toh is software ki help se apni productivity ko double kar sakte hain aur yeh hua bhi. Pehle jahan pe koi kaam karne mein aapko 6 ghante lagte the, ChatGPT ki help se shayad woh kaam ab teen ghante mein kar pa rahe the. So on the whole, hua kya ki is wave of adoption ke baad poori duniya mein logo ne ek productivity boom dekha collectively. Sabhi ka productivity badh gaya, sabhi ka kaam karne ka tarika thoda aasaan ho gaya aur yahan par ChatGPT ne apna true power dikhaya. Aur phir aata hai third wave of adoption which is The API Revolution. Is particular stage mein kya hua ki ChatGPT ne, matlab OpenAI ne sirf ChatGPT ko release nahi kiya tha. Unhone saath hi saath apne GPT models ke APIs ko bhi release kar diya tha general public ke liye. Unka kehna tha ki agar aap chaho toh apne existing software mein ChatGPT jaisi chatting capabilities integrate kar sakte ho. Aur hua bhi yeh. Bahut sari companies ne realize kiya ki although humara software bahut achha hai, but it would be great agar hum kisi tarike se ChatGPT jaisi intelligent features apne software mein la payein. And this was the first time jab Microsoft aage badha aur usne Word, Excel, PowerPoint mein Copilot ka feature add kiya. Which was basically a way to communicate with their software in a natural way. Phir Google aage aaya aur Google ne Gmail, Docs, Sheets, and Drive mein AI ko integrate kiya. And this was the first time jab new AI-first tools aane lag gaye jaise ki Cursor, ya phir Perplexity. Aur hua yeh ki ab AI sirf ChatGPT tak restricted nahi tha. Hum jo softwares pehle se bhi use karte aa rahe the, unme bhi AI capabilities aane lag gayi. Toh in short, hua kya ki is wave of adoption ke baad AI thoda aur zyada accessible ho gaya. Mujhe agar AI ki zarurat hai toh woh sirf ChatGPT mein nahi mil raha, woh mujhe mere aaspaas ke alag-alag softwares mein bhi mil ja raha hai. Toh in teen waves mein aap keh sakte ho ki LLMs humari duniya mein aa gaye completely. So ChatGPT aur uske APIs ke aane ke baad humare aaspaas ke sare softwares AI-enabled ho gaye. Aur iske saath saath yeh bhi make sure ho gaya ki AI bahut zyada accessible ho jaye. But is accessibility ke saath ek nayi problem emerged ki. Aur main is problem ko bulata hoon The Problem of Fragmentation. So mere kehne ka yeh matlab hai ki humare aaspaas bahut sare softwares hain jo hum on a daily basis use karte hain. Aur ChatGPT aur uske APIs ke aane ke baad woh sare ke sare softwares AI-enabled ho gaye. For example, hum Notion use karte hain, hum Slack bhi use karte hain aur yeh dono AI-enabled hain. But the problem is ki Notion ka jo AI hai usko koi idea nahi hai ki Slack ke AI mein kya chal raha hai. Ya phir agar hum VS Code use karte hain aur wahan pe humne coding assistant install kar rakha hai, toh is coding assistant ko koi idea nahi hai ki Microsoft Teams mein kya discussion chal rahi hai. So basically, achanak se humne yeh realize kiya ki hum multiple AI worlds mein jee rahe hain. Ek AI world hai Notion ka, ek AI world hai Slack ka, ek AI world hai humare VS Code ke coding assistant ka, and so on. Aur ek chota sa kaam bhi agar hume karana hai, toh hume in multiple AI worlds ke beech mein juggle karna pad raha hai. Thoda information ek jagah pe padha hua hai. Thoda information dusri jagah pe padha hua hai aur humara kaam hai un sari jagaho pe ja karke us information ko merge karke humara kaam nikalna. Ab soch ke dekho. Hum yeh problem jo face kar rahe hain yeh hum kabhi chahte nahi the. Humara jo vision tha jab humne first time ChatGPT ko dekha aur use kiya. Toh humara jo vision tha woh yeh tha ki kaash hume ek unified AI agent mil jaye jo humare poore kaam ko end to end samajhta ho. Aur not only samajhta ho, jab main kisi problem mein phassu, toh woh unified AI agent meri koi bhi problem ko solve bhi kar sake. This was our vision. But instead hume kya mila? Hume mile paanch alag-alag AI tools. Thoda kaam is AI tool se nikalwa lo, thoda kaam dusre AI tool se nikalwa lo, and so on. Ab unified AI agent jo hume banana tha, yeh banane ka kaam actually aasaan nahi hai. Aur is unified AI agent ko banane ke raste mein jo sabse badi problem hai woh problem hai problem of context, jo hum next samjhenge. Ab baat karte hain context ki. Context MCP mein ek bahut fundamental concept hai. In fact, itna fundamental hai ki MCP ke naam mein bhi context aata hai, Model Context Protocol. Toh ek kaam karte hain. Thoda sa time le karke is concept ko detail mein samjhte hain. Sabse pehle, ekdum simple shabdon mein samjhte hain ki context hota kya hai. So, in simplest words, context is everything an AI can see when it generates a response. AI jab koi response generate karta hai, toh us response ko generate karte time usko jitni bhi cheezein dikhai de rahi hain, usi ko hum context bulate hain.
[13:05]Thoda formal definition yeh hoga ki Context refers to the information (conversation history, external docs etc) that the LLM uses to generate a response. Ab woh information conversation history bhi ho sakta hai, external documents bhi ho sakte hain.
[13:20]Ek example ke through samjhte hain. Maan lo aap ChatGPT se baat kar rahe ho about Quantum Physics. Aapne poocha Quantum Physics kya hota hai, Quantum Physics kahan se seekhein, kaun si achhi books hain. Ab suddenly se aapne poocha ki yeh particular topic seekhna kitna difficult hai. Ab ChatGPT ko kaise pata kis baare mein baat ho rahi hai? The answer is conversation history. ChatGPT jhat se jaa ke poora conversation history dekhega aur woh realize karega ki user Quantum Physics ke baare mein baat kar raha hai. Toh woh samajh jayega ki woh Quantum Physics kitna difficult hai padhna yeh pooch raha hai aur woh jhat se aapko answer dega. Toh is particular scenario mein jo aapka conversation history hai, wahi aapke AI model ka context hai. theek hai? Agla response type karte time aapka AI us poore conversation history ko dekh sakta hai aur uski help se apna response print kar sakta hai. theek hai?
[14:21]Toh I hope aapko context roughly samajh mein aa gaya. Ab is particular example mein context ek bahut simple tarike se dikhaya gaya hai. Basically ek conversation history ke form mein. But context zaruri nahi hai ki itna simple ho. Especially agar aap professional use cases ki baat karo. Toh ab main aapko ek dusra example leke samjhata hoon ki context kitna difficult ho sakta hai samajhna. So hum baat karte hain ek software engineer ki. Aur uske life mein ek particular din kya chal raha hai uske baare mein baat karte hain aur wahan pe samjhte hain ki context kya hai. theek hai? Toh maan lo main ek software developer hoon aur main ek startup mein kaam karta hoon. Humara koi product hai. Maan lo EdTech product hai jahan pe log courses purchase karte hain, just like Udemy. Ab suddenly se mere ko ek requirement di gayi aur mujhe bola gaya ki mujhe humari website mein two-factor authentication ka feature add karna hai so that humari website thodi zyada secure ho jaye. Toh yeh poori cheez hoti kaise hai woh main aapko batata hoon. So sabse pehle kya hoga, Jira jaise kisi software pe ek ticket raise kiya jayega. Aur woh ticket mujhe assign ho jayega ki mujhe yeh feature develop karna hai. Ab yeh feature develop karne ke liye main kya karunga, sabse pehle GitHub pe jaunga aur wahan se sabse updated code base ko apne machine pe pull karunga, right? Aur phir obviously two-factor authentication karna hai toh uske liye mujhe database mein ja karke samajhna padega ki humara existing schema kaisa dikhai deta hai. Toh main MySQL jaisa koi software use karunga jo humari company use kar rahi hai aur us software mein ghuss kar ke main database ka schema study karunga. Suddenly se mujhe yaad aayega ki mujhe certain security guidelines bhi follow karne hain in order to implement the two-factor authentication. Toh main kya karunga main Google Drive pe jaunga aur wahan se main ek security document utha ke launga jiski help se main yeh poora two-factor authentication implement karunga. Finally, agar main kahin pe phass jaunga, toh main ja karke apne team-mate se baat karunga ya help lunga with the help of Slack jaisa koi software. Toh is particular case mein if you can see, mujhe jo kaam karna hai, two-factor authentication system develop karna hai, yeh mera task hai. But is task se related jo context hai woh kai jagahon pe exist kar raha hai. Ab yeh itna simple nahi hai jaise humare last example mein sirf ek conversation history mein humara poora context exist kar raha tha. Yahan par is scenario mein humara context multiple jagahon pe saath mein exist kar raha hai. Ab khud soch ke dekho, agar mujhe yeh two-factor authentication develop karne wala kaam karna hai with the help of AI, let's say ChatGPT, toh main kaise approach karunga? Sabse pehle main Jira pe jaunga, wahan se ticket mein jo bhi likha hua hai ki kaise-kaise implement karna hai is poore solution ko woh main copy karunga, ChatGPT pe paste karunga. Uske baad main code base mein jaunga, 10-12 files ka code copy karunga jo humara existing authentication system hai, uska code paste karunga ChatGPT mein. Uske baad main MySQL pe jaunga, schema copy karunga, ChatGPT pe paste karunga. Uske baad main apna security document fetch karke launga, wahan pe jo specifications likhe hue hain, usko ChatGPT pe paste karunga. Uske baad Slack mein agar koi discussions hue hain, main usko copy karunga, ChatGPT pe paste karunga. Aur finally, itna kuch paste karne ke baad, 20 minute ke baad, main apna pehla question poochunga, How to implement two-factor authentication in such a system?
[18:10]Ab I guess aapko dikhai de raha hai ki context jo hai woh kitna scattered hai aur context ko banana kitna mushkil kaam hai aur kitna time-taking kaam hai. Aur yahi sabse badi problem hai. Humari problem yeh hai ki ek simple question poochne ke pehle hume hazaaron lines of code copy karke pehle ChatGPT pe dalna pad raha hai. Matlab, in a way agar aap bolo toh hum developers ek tarike se human APIs ban gaye hain. Aur humara kaam kya hai? Humara kaam hai ChatGPT jaise software ke liye context ko assemble karna.
[18:50]So, we are basically human APIs. Aur in a way yeh kehna galat nahi hoga ki yeh jo context assembly hum kar rahe hain, isme hum zyada time spend kar rahe hain rather than actually developing the product. Aur iske alawa aapko poore time is baat ka khayal rakhna hai ki aapne kya context provide kiya hai AI ko, kya bacha hua hai, kya AI bhool gaya, kya summarize karke batana hai.
[19:15]Yeh poori cheez bilkul bhi scalable nahi hai. Soch ke dekho agar aapke paas ek aisi company ka code base hai jahan pe 50,000 lines ka code, code base hai. Toh kya aap usko summarize kar sakte ho apne AI ke liye, ya phir aap kya woh poora context copy-paste kar sakte ho ChatGPT mein? Toh it's not at all possible ki aap is tarike se alag-alag jagaho se context assemble karo aur apne AI ko do. And that is why I am saying ki ek unified AI agent banana bahut mushkil hai because jo sabse bada challenge humare saamne hai woh hai context assembly ka. Ek real professional scenario mein aapka context across systems exist karta hai. Aur un sari cheezo ko copy-paste-paste karke ek jagah pe dal pana aur phir ChatGPT se baat kar pana ya kisi bhi AI software se baat kar pana is a very-very difficult thing to do and at scale it's not at all possible. So agar main is context wali problem ko summarize karun toh main problem yeh hai ki agar hum yeh chahte hain ki humara AI chatbot humare kaam se related sari cheezo ko samajh paye aur problems ko solve karne mein hume help kar paye, toh it is important ki woh humara sara kaam dekh paye. Basically, humara sara kaam uske context ka part hona chahiye. But problem yeh hai ki humara jo context hai, humara jo kaam hai, woh scattered hai. Toh phir AI ko woh sara kaam dikhana ek bahut mehnat ka kaam ban gaya because hum manually copy-paste kar rahe hain cheezo ko. Aur jaise-jaise aapka project bada hota jayega, jitne zyada tools ke saath aap kaam karoge, yeh copy-paste karne wali cheez fail karti chali jayegi. Ideally, achha hota agar kisi tarike se ChatGPT jaisa software khud-ba-khud sari jagaho se jaa karke necessary context fetch karke la pata. Toh phir yeh jo manual copy-paste wala poora kaam hai yeh hume nahi karna padta. And guess what, yeh problem solve ki gayi eventually. So hua kya? OpenAI ne mid of 2023 ek naya concept release kiya, jiska naam tha Function Calling. Ab yeh ek bahut hi simple but bahut hi powerful concept tha.
[21:35]Function calling ki help se aap kya kar sakte ho ki aap apne LLM se kisi external function ko call karwa sakte ho. Which basically means ki ab aapka LLM sirf chat karne ke hi kaam nahi aayega. But agar zarurat padi, toh woh koi task bhi complete kar sakta hai. So function calling mein hota kya hai ki aap apne LLM ko kuch set of functions provide karte ho. Aur aap apne har function ke baare mein ek description bhi LLM ko de dete ho ki yeh particular function yeh kaam karta hai. theek hai? Ab maan lo humne apne LLM ko ek function de diya, load file bol ke. Aur humne yeh description de diya ki agar kabhi future mein user chahega kisi file ka content load karna, toh you can use this function. Toh ab hoga kya ki jab user agar aisa prompt deta hai, Read the content of the file abc.txt. Toh LLM normal chatting karne ke badle yeh samajh jayega ki user usko kuch kaam karne ko bol raha hai. Toh woh jhat se kya karega list of functions ko ja karke scan karega aur jo bhi function ka description match karta hai given task se, automatically LLM kya karega us function ko call karne ko bolega with the right set of arguments. Jaise yahan par woh bolega ki load file ko call karo with this argument abc.txt aur phir hum kya karenge is load file function ko call karenge is input ke saath aur humara task execute ho jayega. Ab hai toh yeh bahut chota sa concept, but yeh path-breaking tha. Because for the first time, ab aap LLM se sirf chat nahi kar sakte, but saath hi saath aap koi task bhi execute kara sakte ho. So phir kuch is tarah ke architecture picture mein aane lag gaye ki ek LLM hai jisse humne bahut sare functions ko connect kar rakha hai, bahut sare tools ko connect kar rakha hai. Ek tool hai weather API se weather data fetch karke laane ke liye, ek tool hai database mein queries run karne ke liye, ek dusra tool hai GitHub se aapke repository ka information fetch karke laane ke liye. Ek aur tool hai web search karne ke liye. So basically, aap apne alag-alag task ke liye alag-alag functions create kar sakte ho aur har function internally yeh batayega ki us given task ko execute kaise karna hai. Aur aapke LLM ka sirf ek kaam hai ki user ke prompt ko woh read karke samjhega ki kaun se tool ko invoke karna hai. Aur yeh ek bahut hi revolutionary idea tha. Iske aane ke baad tools ki toh jaise bauchar ho gayi. Koi bhi company jo already LLMs ke saath kaam kar rahi thi, unhone power realize kiya tools ka aur unhone apne workflow ko seamless banane ke liye, context assembly ko seamless banane ke liye bahut tarah ke tools banana shuru kar diya. Jaise ki sabse pehle unhone enterprise softwares ke liye tools banaye. Jaise har company aapka Salesforce use karti hai, Slack use karti hai, Google Drive use karti hai, databases use karti hai. Toh sabse pehle companies ne apne developers ko bola ki aap ek kaam karo, aap humare AI chatbot ke liye Salesforce Salesforce ka integration ka ek tool banao. Similarly aap Slack ke liye ek Slackbot tool banao. Aap Google Drive ke connectors bana lo, aap database pe query run karne ke tools bana lo, ya phir aap GitHub ka integration ka tool bana lo so that yeh jo sari jagaho pe humara context exist karta hai, inse hum connect karke kisi bhi time pe context ko fetch karke la payein. Phir companies ne internal tools bhi banaye jaise HR departments ne employees ka data access karne ke liye alag functions bana liye, tools bana liye. Finance team ne accounting systems ke liye tools bana liye. Marketing team ne campaign management ke liye tools bana liye, IT department ne infrastructure management ke liye tools bana liye aur isi point pe kya hua, jo naye AI-first softwares the, unhone bhi bahut tarah ke tools banana shuru kar diya. Jaise ki Cursor jo humara AI-powered code editor hai usne file system access ka tool banaya jisse aap apne local file system pe koi bhi file ko access kar sakte ho aur usko intelligently search kar sakte ho. Similarly Perplexity jo hai usne web browsing and real-time information retrieval ka tool bana liya. ChatGPT Plus bol karke ek subscription-based feature aaya ChatGPT mein jiski help se aap browsing kar sakte ho, files upload kar sakte ho, codes execute kar sakte ho. Claude ne computer use bol ke ek feature launch kiya jiski help se woh aapke computer ko completely control kar sakte hain. So basically, function calling ke aane ke within six months mein poore AI world mein tools ki bauchar si ho gayi. theek hai? Aur ab dekhte hain ki yeh tools ke aa jaane se humara yeh jo software wala scenario tha jahan pe context scattered tha, woh scenario ab kaise pan out kar raha hai. So abhi bhi main hi woh developer hoon aur mujhe woh same kaam diya gaya ki mujhe two-factor authentication implement karna hai. Aur yeh kaam karne ke liye main ChatGPT ki help chahta hoon. But ab mere paas sare ke sare necessary tools hain. Jaise ki jaise hi Jira pe ek ticket raise hua aur mujhe assign kiya gaya toh now what I can do is rather than manually Jira pe jaa ke ticket ka content copy-paste karne ke maine ChatGPT ko seedhe bola ki go and check whether I have a new ticket assigned on Jira or not. Aur since ChatGPT ek tool ke through Jira se connected hai toh yeh sara ka sara kaam behind the scenes automated tarike se ho gaya. Phir mujhe pata chala ki haan mujhe ek naya ticket assign kiya gaya hai. I have to develop two-factor authentication. Toh maine jhat se code ChatGPT ko bola ki can you fetch the most updated code base from GitHub? Toh again ek tool ke through GitHub connected hai toh sara ka sara code base aa gaya. Phir maine ChatGPT ko bola ki yaar mujhe two-factor authentication banane ke liye schema chahiye hoga. Toh since main MySQL se connected hoon with the help of a tool, jhat se schema bhi ChatGPT ke paas aa gaya. Phir maine bola security guidelines fetch karke le aao Drive se. Drive ke liye bhi connector exist kar raha tha. Aur finally maine poocha ki mere team wale kya baat kar rahe hain Slack pe is particular cheez ke baare mein woh information bhi fetch karke le aao. Ab yeh sara ka sara context assemble ho gaya. Ab maine apna question poocha ki now that you can see everything, tell me how can I build a two-factor authentication system. Now you can see the power of tools ki ab jo context pehle scattered tha, woh ab connected hai humare AI chatbot se, ChatGPT se. And now ChatGPT can actually see my entire work. Aur since woh mera poora kaam dekh sakta hai, toh obviously us poore kaam ko karne mein mujhe bahut sahi tarike se help bhi kar sakta hai. So is tarike se humne woh context assembly wali jo problem thi jo hume rok rahi thi ek full-fledged unified AI partner banane se us problem ko solve kiya tools ke aane se. Ab jab yeh function calling/tools wala solution aaya toh logo ko aisa laga ki unhone context assembly wala problem solve kar liya hai kuch dino tak at least. But kuch time ke baad unhone realize kiya ki bhale hi yeh tools wala solution kaam karta hai aur sara ka sara context aapko la karke deta hai, but is approach mein ek bahut bada flaw hai.
[29:12]Flaw kya hai main aapko samjhata hoon. So jaisa maine aapko bola aapko apne AI ko context provide karna hai, context scattered hai across different tools. So aap kya karte ho, har tool ke liye ek function likh lete ho. Jaise yeh code main aapko dikhata hoon. Yeh humare LangChain playlist mein hum jo calculator bana rahe hain, I am sorry, jo chatbot bana rahe hain uska code hai aur us chatbot mein humne do tools add kiye hain. Ek calculator ka tool aur ek stock market mein kisi bhi company ka stock value fetch karke laane ka tool aur aap dekho in dono tools se associated humne ek-ek function bana rakha hai. This is one function. This is another function. So basic funda yeh hai ki aapko apne har tool ke liye ek function ka code likhna padta hai. Ab problematic kaise hai yeh cheez main aapko batata hoon. So maan lo ek scenario aisa hai ki hum jis company mein kaam karte hain, wahan pe hum teen alag-alag AI chatbots ke saath kaam karte hain. Ek hai normal chatbot jisse hum company related kuch bhi pooch sakte hain. Dusra hai coding agent. So humari company mein jitne bhi developers hain, software developers, woh yeh coding agent use karte hain. Aur ek aur chatbot bhi hai, analytics agent bol ke. So humare jitne bhi data analyst hain, unko hum yeh wala chatbot dete hain data analysis karne ke liye, automated data analysis karne ke liye. Ab thodi der ke liye maan lo ki humari jo company hai woh kuch 20 alag-alag SaaS tools ke saath kaam karti hai. Jaise ki Jira, Slack, GitHub, MySQL, Drive aur is tarah ke aur 15 tools. Toh ab khud soch ke dekho ki hume kitne functions likhne padenge. Hume is tarah ke kitne functions likhne padenge. I guess aap samajh pa rahe ho ki agar aapke paas N AI chatbots hain aur aapke paas M different tools hain ya services hain, toh aapko in total N cross M integrations ya functions code karne padenge. Ab yeh toh chota number hai, badi companies mein yeh numbers aur bade ho sakte hain. Ho sakta hai aap aur zyada chatbots ke saath kaam karte ho ya phir aap aur zyada SaaS tools ya internal tools ke saath kaam karte ho. Now making so many functions, writing so many functions, coding so many functions is a development nightmare. Soch ke dekho. Because aapko har function ke liye apna khud ka authentication method likhna padega. Har tool ka apna khud ka data format hoga, API patterns honge. Har tool alag-alag tarike se error handling mangega. Toh basically itna code karna khud mein ek alag task hai. Aapko basically ek alag software team bithaani padegi just to create these integrations. theek hai? Iske alawa aur bhi problems hain. Problem hai maintenance ki. Maan lo aapke paas teen chatbots hain, 20 integrations hain. Total kitne aapko functions likhne pade? 60 functions. Ab aapko in 60 functions ko maintain bhi toh karna hai. Kal ko agar Google Drive ne jaa karke apne API mein thode se changes kar diye, toh automatically aapke jitne bhi Google Drive wale integrations hain woh toh sare fail kar gaye. Aapke tino charo chatbots mein ab Google Drive se baat hi nahi ho pa rahi. Toh phir se aapko ja karke usko debug karna padega, right? Phir iske alawa security ki problem hai. Aap jab badi companies mein kaam karte ho toh you make sure ki aapka jo software hai woh secure tarike se yeh sara ka sara kaam kare, but since aapne 60 different integrations likh rakhe hain aur sab mein khud ka apna auth hai, apne API keys hain. Toh in sabko ek single jagah se aap manage nahi kar sakte, alag-alag files mein yeh sara information pada hua hai. Toh security aapki fragmented hai aur wahan pe ho sakta hai ki kisi tarike se koi hack perform ho jaye. And the fourth problem is ki is poore approach mein aapka cost and time, dono waste ho raha hai. Soch ke dekho, time kaise waste ho raha hai. Agar mere paas ek chatbot hai aur us us mujhe uske saath 20 tools ko connect karna hai, toh mujhe 20 tools ke liye individual integrations likhne pad rahe hain. Mujhe basically is tarah ke 20 functions banane pad rahe hain jo ki khud mein bahut complex baat hai. Toh har integration, har function ko code karne mein time lagega. Toh mera chatbot fully functional ho, isme ho sakta hai do mahine, teen mahine ka time lage, right? Uske baad cost involved hai. Mujhe basically ek poori team setup karni pad rahi hai jo poora ka poora yeh integration wala part hai, usko handle kare. Ab unki salaries mujhe dekhni pad rahi hai. Ab soch ke dekho, mera end goal kya tha main yeh poori cheez kyon kar raha tha? So that mere jo existing developers hain unka kaam aasaan ho jaye. Ab unka kaam aasaan ho, is kaam ko karne ke liye mujhe dusre developers hire karne pad rahe hain, just see the irony in this. Aap koi cheez ko aasaan karne gaye the, but woh ulta aur mushkil ho gayi. Toh what I am trying to say is ki yeh jo tools aur function calling wala jo solution humne implement kiya in order to solve the context assembly problem, woh actually at scale khud mein ek integration problem leke aaya jisko solve karna again logo ko bahut bhaari lagne laga. Ab hum log ek baar is function calling/tools ka jo integration problem hai, usko summarize karte hain. Basic overview yeh hai is problem ka ki humari company ek se zyada AI chatbots use kar rahi hai. Ho sakta hai ki humari company Perplexity use kar rahi ho, saath hi saath Cursor use kar rahi ho aur ChatGPT bhi use kar rahi ho. Aur hum chahte hain ki yeh teenon chatbots GitHub se connect karein. Ab problem yeh hai ki yeh kaam karwane ke liye hume teen alag integrations banane padenge. One for Perplexity and GitHub, one for Cursor and GitHub and one for ChatGPT and GitHub. Aur problem yeh hai ki yeh teenon integrations ek doosre se alag-alag dikhenge. So the problem is that every AI tool is building its own way to call every API. Achha hota agar hum kisi tarike se yeh make sure karte ki rather than making three integrations, hum sirf ek single integration banate GitHub ka. Aur wahi integration Perplexity ke saath bhi kaam kar jata, wahi Cursor ke saath bhi kaam kar jata, wahi ChatGPT ke saath bhi kaam kar jata. This is what we want to achieve and this is how we can solve this integration problem. Aur phir isi problem ko solve karne kaun aaya? MCP, Model Context Protocol. Toh chalo guys, ab baat karte hain MCP ki. Main aapko ekdum simple shabdon mein batata hoon ki MCP kaam kaise karta hai. So MCP mein do entities hoti hain. Ek hota hai client aur dusra hota hai server. Aur poora ka poora communication inhi do entities ke beech mein chal raha hota hai. theek hai? Ab aap yeh mat sochne lag jana ki yaar yeh kya nayi cheez aa gayi client bol ke ya kya nayi cheez aa gayi server bol ke. Actually, MCP mein jo client hota hai na woh aapka AI chatbot hi hota hai. theek hai? Basically aapka ChatGPT hua, Cursor hua, Perplexity hua, ya phir aapka khud ka AI chatbot, usi ko hum client bula rahe hain. Aur server jo hota hai woh basically woh service hoti hai jisse aap apne AI chatbot ko connect karwana chahte ho. Jaise ki GitHub ya Google Drive ya Slack, theek hai? Ab jo general architecture rehta hai kisi bhi MCP application ka woh kuch aisa dikhai deta hai ki aapke paas ek single client hota hai aur woh bahut sare servers se connected hota hai, bahut sare tools se connected hota hai. Matlab ek AI chatbot multiple tools ke saath connected hai. Toh ek tool ho gaya GitHub, ek ho gaya Google Drive, ek ho gaya Slack. Ab yeh jo poora ka poora communication ho raha hota hai, poori ki poori jo baat cheet ho rahi hoti hai between the client and the servers. Yeh jis language mein ho rahi hoti hai na, usi ko hum MCP ya Model Context Protocol bulate hain. theek hai? Toh I really hope aapko abhi tak jo bhi maine bola, woh samajh mein aa raha hai. Ab I know aapke mann mein yeh curiosity aa rahi hogi ki kaise main apne AI chatbot ko MCP client bana sakta hoon aur kaise main apne tools ko MCP server bana sakta hoon. Ab iska answer yeh hai ki aap agar MCP protocol ko properly padho, uska documentation properly padho, toh yeh poora code aap khud se scratch se bhi likh sakte ho.
[38:21]Jo yeh MCP wala communication establish karega between client and server. But Anthropic aapko ek bani banayi library ya SDK deta hai. So agar aap apne AI chatbot ko MCP compatible client banana chahte ho toh aapko kuch nahi karna hai, ek MCP client SDK aapko install karna hoga apne machine pe jahan pe bhi aapka AI chatbot chal raha hai.
[38:47]Aur agar aap apne tools ko MCP compliant server banana chahte ho toh again aapke paas ek library hai, MCP server SDK bol ke, aap isko use karoge to build your tools. theek hai? This is the biggest difference.
[39:08]Isi se aapke MCP client ban rahe hain, isi se aapke MCP server ban rahe hain. Ab ek baar thode technical level pe ja karke samjhte hain ki MCP kaise different hai function ya tool calling se jo humne just iske pehle padha. So function ya tool calling mein aap kya karte ho ki aapko jis bhi tool se connect karna hai, aap uski API ko access karne ke liye ek function likhte ho apne AI chatbot ke andar jo main aapko thodi der pehle bhi dikhaya bhi tha. Humare chatbot mein humne do alag tools banaye the, ek calculator wala, ek stocks wala. So hota kya hai ki yahan pe bhi kind of ek client server model follow ho raha hota hai. Jaise ki maan lo agar hum apne AI chatbot ko ek weather tool ke saath connect karna chahte hain. Toh obviously kahin na kahin server pe weather tool ka ek API hoga jo shayad fast API mein likha hua hai aur kuch aisa dikhai dega. Toh yeh kind of humara server ho gaya aur phir isko access karne ke liye hum kya karte hain ki hum ek function likhte hain, humare AI chatbot ke code mein. theek hai? Aur yahan pe hum code likhte hain is API ko hit karke data laane ka. Toh yeh kind of humara client side ho gaya. Toh yeh hai normal function ya tool calling. Ab main aapko batata hoon ki MCP mein kya change hota hai. So MCP mein bhi ek server hoga, weather wala jahan pe weather data recorded hoga. theek hai?
[40:35]Bas yahan pe jo server banega yeh MCP library ki help se banega. Thoda sa code alag hai but internally it is using the same API. theek hai? The main difference comes from the client side.
[40:50]Client ke side pe aapko apne AI chatbot mein koi code likhne ki zarurat hi nahi hai. Since aapne pehle se client aur server ko integrate kar diya hai, configure kar diya hai and they are speaking the same language, that is Model Context Protocol, toh aapko alag se code likh karke is API ka data fetch karne ki zarurat nahi hai. Yeh sara ka sara kaam server handle karta hai. Toh main ek baar phir se aapko technical level pe samjhata hoon. Function calling mein ek server ka code hai jo company ne likh rakha hai. Aur ek client ka code hai jo aapne likh rakha hai aur yeh dono code saath mein kaam karte hain tab poora ka poora task completion hota hai. MCP mein sara ka sara heavy lifting server kar raha hai, MCP server.
[41:48]Aur client side pe AI chatbot wali side pe aapko kuch bhi nahi karna bas connection karke baith jana hai. theek hai? So maan lo agar aap ek GitHub server bana rahe ho MCP mein, toh yeh GitHub server ke upar har cheez karne ki responsibility hai.
[42:25]Poora ka poora business logic is server pe implemented hoga. Authentication ka kaam server pe implemented hoga. API ka rate limiting server dekhega. Data format translation LLM ko saamne kya samajh mein aa raha hai yeh sab server dekhega. Error handling ka code bhi server pe hi likha hoga aur woh sahi error codes client ke paas bhejaega agar kuch gadbad hoti hai aur aapke client ko kya karna hai, simply connect kar jana hai is server se aur usko same language bolni hai, MCP. This is the biggest difference between function/tool calling and MCP. I really hope aapko thoda sa samajh mein aa raha hai. Aage ke videos mein hum isko bahut detail mein study karenge, but abhi main aapko ek overview, ek gist dena chahta tha. Ab ek baar discuss karte hain ki is particular logic ko implement karne se is particular tarike ko implement karne se kya benefit aata hai. So jaisa maine bola ki sabse bada benefit MCP ka yahi hai ki yahan pe aapko client side pe kuch bhi nahi karna. Sara ka sara heavy lifting servers kar rahe hain. Toh MCP ke aane ke baad yahi hua, jitne bhi famous service providers the, GitHub hua, Google Drive hua, Slack hua, in sab ne apne-apne official MCP servers bana liye.
[43:53]Which basically mean ki koi bhi MCP compliant AI tool can easily connect with these servers. Ab isse fayda yeh ho raha hai ki aapki jo integration problem thi, maan lo aapki company mein teen chatbots hain aur aap 10 services ke saath connect karna chahte ho toh aapko 3 multiplied by 10, 30 unique integrations likhne pad rahe the. Ab aapko aisa nahi karna hai. Ab aapko agar 10 MCP server se connect karna hai toh aapko sirf 10 integrations chahiye aur woh integrations bhi server ki side pe likhe ja rahe hain jo service provider hai woh khud likh raha hai. Client wali side pe aapko koi bhi integration likhne ki zarurat hi nahi hai. Aap seedhe bas configure kar sakte ho apne AI tool ko to connect with the server. So pehle jahan pe M cross N integrations the, ab uske badle sirf M + N integrations hain. Aur usme bhi jo actual code likhne ka kaam hai woh server ki side pe delegate ho gaya, client ko kuch nahi karna. This is the first benefit. Second benefit is there is no maintenance overhead. Main apni side pe since koi code likh hi nahi raha connection karne ka toh phir wahan pe kuch gadbad ho hi nahi sakti. Kal ko agar update hota hai API, toh server ka headache hai woh. Server apna code update karega, mujhe apni side pe kuch bhi karne ki zarurat nahi hai. Client side pe there will be no changes. Third, reduced cost and time. Maan lo mujhe apne AI chatbot ko 10 alag services ke saath connect karna hai. Toh pehle mujhe woh 10 integrations banane padte the jisme time lagta tha, but aaj ki date mein mujhe kya karna hai, mujhe simply AI chatbot banana hai aur 10 servers already available hain, main unse directly connect kar jaunga on day one. So day one pe hi mujhe 10 servers ka jo bhi kaam hai, woh aasani se mil jayega. So yahan pe time bach raha hai aur obviously cost bach raha hai because mujhe alag se engineers ko hire nahi karna to create and maintain these integrations. And the last point is ki we get better security. Why? Because jab aap ek AI chatbot ko connect karte ho to multiple services, toh wahan pe aapko simply ek JSON file maintain karni hai, ek config file maintain karni hai jahan pe aapke sare ke sare connections ek jagah pe likhe hue hain.
[46:14]Jaise ki for example itna part jo aapko dikhai de raha hai yeh mere AI chatbot ko GitHub se connect karne ka code hai. This is it. Bas itna hi code mujhe chahiye. Yahan par maine apna personal access token wagerah de diya hai aur yeh ab mere GitHub account se connect ho gaya. Similarly, sirf itna code mere AI chatbot ko Notion se connect kara raha hai. Again maine apna yahan pe API key wagerah de diya hai. theek hai? Ab is ek single file ko manage karna ya audit karna is much simpler than purane wala case jahan pe humne 10 alag files mein 10 alag tools mein alag-alag API keys rakhe the, right? Toh this is a much better security model as well. So these are the main benefits of MCP. Again, sab kuch sirf ek simple fact ke upar depend kar raha hai aur woh fact yeh hai ki MCP mein the entire heavy lifting is done by the server.
[47:24]On the client side, the AI chatbot has to do nothing. You basically have to maintain a JSON file like this to connect to various servers, aur that's it. Aapka kaam ho gaya. So MCP kaise kaam karta hai uska ek overview toh main aapko de diya. But ab main aapke saath ek aur cheez discuss karna chahta hoon. Aur woh yeh hai ki MCP kaise itni tezi se popular ho raha hai. Kya reason hai ki MCP ka ecosystem itni tezi se grow kar raha hai? Aur kyon aisa lag raha hai ki MCP agle teen se paanch saalon mein industry standard ban jayega. Ab uske piche reason bahut simple hai. Yeh poora dynamics aap bahut aasani se samajh sakte ho. So ho kya raha hai ki kuch bahut famous AI chatbots openly bol rahe hain ki hum MCP ko support karte hain. Claude Desktop ne bola ki we support MCP, Cursor, Windsurf, Perplexity ne bhi bola we support MCP. Ab jaise hi in chatbots ne bola ki hum MCP ko support karte hain, toh bahut sare jo services hain unke upar pressure aane lag gaya. Services like GitHub, Slack, Google Drive, etc. Pressure isliye aane lag gaya because in companies ko dikhai deta hai ki future mein jo bhi log kaam karenge woh sara kaam in AI tools ke through karenge. So agar yeh AI tools MCP support kar rahe hain, toh mujhe bhi apne APIs se MCP servers bana lene chahiye.
[49:13]Kyonki agar maine ek baar MCP server bana liya toh koi bhi MCP compliant compatible tool easily mere server se connect kar sakta hai. Usko koi custom code likhne ki zarurat nahi hai.
[49:59]Toh basic funda yeh hai, zyada AI clients aa rahe hain client ban ke toh bahut sare servers bhi ban rahe hain. Bahut sare servers ban rahe hain toh naye chatbots ke upar yeh pressure aa raha hai ki unko connect karna chahiye with these servers.
[50:24]Toh naye wale jo clients ban rahe hain woh automatically MCP compliant ho gaye. theek hai? Toh basically ek network effect sa ban raha hai jiski wajah se yeh ecosystem exponentially tezi se grow kar raha hai. So zyada adoption, zyada standardization, zyada ecosystem ki value. Aur agar koi client ya phir koi server basically koi AI chatbot ya phir koi tool agar MCP adopt nahi kar raha hai, toh phir uska yeh matlab hai ki woh yeh jo massive ecosystem ban raha hai usse cut-off ho jayega. Aur usko bahut sara custom code likhna padega in order to make sure ki uska AI chatbot sahi se sare services ke saath interact kare, jo ki again bewakoofi hogi. That is why MCP ekdum sahi tarike se khud ko position karke baitha hai aur aisa lag raha hai ki agle teen se paanch saalon mein it will become an industry standard. theek hai? Toh with that, I'll conclude this video. I really hope main aapko ek story ke tarike se samjha paya ki MCP ki zarurat kyon hai. Maine ekdum day one se shuru kiya aur aaj jo scenario hai wahan tak main aapko leke aaya. I really hope video lamba hua but boring nahi laga aur aap kuch value derive kar paye. Ab agle video mein hum bahut detail mein architecture level pe ja karke samjhenge ki MCP kaam kaise karta hai, The What of MCP. Toh I really hope you are excited for the next video. Agar aapko yeh video pasand aaya, please like karna. Agar aapne is channel ko subscribe nahi kiya hai, please do subscribe. Milte hain next video mein.
