YouTube2Text

MCP In 26 Minutes (Model Context Protocol) — Transcript

by Tina Huang · 5,477 words · 787 segments · language en · Watch on YouTube

Full transcript

  1. 0:00I learned MCP for you. In the past 68
  2. 0:02months, you've probably heard a lot of
  3. 0:04people talk about MCP. MCP is so
  4. 0:06amazing. MCP is such a game changer.
  5. 0:09Well, I must tell you, they are correct.
  6. 0:11I have taken a bunch of courses, build
  7. 0:14MCP servers and clients for our team and
  8. 0:16a couple B2B clients that we work with
  9. 0:18as well. We've also included a section
  10. 0:20on MCP in our AI agents boot camp. So,
  11. 0:22in this video, I'm going to explain to
  12. 0:24you the fundamentals of MCP, show you
  13. 0:26how to use MCP servers and how to build
  14. 0:28your own MCP servers as well, both with
  15. 0:30code and no code. Throughout this video,
  16. 0:32there will be little assessments, which
  17. 0:33if you can answer these questions, then
  18. 0:35congratulations, you'll be educated on
  19. 0:37MCP. Now, without further ado, let's go.
  20. 0:39A portion of this video is sponsored by
  21. 0:41Balt. Okay, a quick overview of the
  22. 0:43structure of today's video. First, we're
  23. 0:44going to define MCP, talk about why it's
  24. 0:46important, and how it is that we use
  25. 0:48them. Then we'll be diving under the
  26. 0:50hood in which I'll cover the
  27. 0:51fundamentals of MCP, the concepts and
  28. 0:53frameworks that govern it. Basically,
  29. 0:55everything you need to know to be able
  30. 0:56to use and build your own servers. But
  31. 0:58it is not enough just to know all these
  32. 0:59concepts. So I'll actually show you how
  33. 1:01to build one as well, both with code and
  34. 1:03no code. Let us now define MCP. MCP,
  35. 1:06which is the abbreviation of model
  36. 1:08context protocol, is defined by
  37. 1:10Enthropic, the people who made it, as an
  38. 1:12open protocol that standardizes how your
  39. 1:13LM applications connect to and work with
  40. 1:15your tools and data sources. Okay. What?
  41. 1:18So first time I read that definition I
  42. 1:20was like sounds important. So let me
  43. 1:23actually explain it to you the way that
  44. 1:24I understood it using an analogy. MCP is
  45. 1:27like this USB plug here. It allows any
  46. 1:30device to be able to connect to any
  47. 1:32other devices to be able to work
  48. 1:33together. Prior to this universal
  49. 1:35standard of the USB port, if you wanted
  50. 1:38to connect things to each other, like
  51. 1:39say your computer to your microphone, to
  52. 1:41your webcam, you would be having a very
  53. 1:43bad time because all of the ports to
  54. 1:45connect with each other would be
  55. 1:47different, which means you would have to
  56. 1:48get a bunch of different special cords
  57. 1:50and different adapters to manage to get
  58. 1:51them to work together. And then you
  59. 1:52would have to write the software to
  60. 1:54cater towards like those specific port
  61. 1:55thingies, etc. Very complex, very messy.
  62. 1:58But with this USB port, we're able to
  63. 2:00now just connect all these devices to
  64. 2:02each other and they just work. Amazing.
  65. 2:04Wow. Do not be fooled. It sounds like
  66. 2:06such a simple little thing, right? Just
  67. 2:07like standardizing the port connectors,
  68. 2:10but it amounted to massive amounts of
  69. 2:11innovation with this ease of use. Now
  70. 2:13before MCP existed, if you have your
  71. 2:16agent and you wanted to give it access
  72. 2:17to different tools and different context
  73. 2:19or different types of data, you would
  74. 2:20have actually faced a very similar
  75. 2:21problem to the pre- USB port days
  76. 2:24because each external system and tool
  77. 2:25that you're trying to connect to has
  78. 2:26their own specific definition of how it
  79. 2:28is that you interact with it, which is
  80. 2:30their special API. For example, if
  81. 2:32you're trying to make a scheduling
  82. 2:33agent, you would need to give it access
  83. 2:34to like your calendar, your email, a
  84. 2:37note-taking app, Zoom, Calendarly, and
  85. 2:39you would need to custom write the code
  86. 2:41for each specific tool that you want to
  87. 2:43give your agent, which is a pain in the
  88. 2:45ass. So that's why when Enthropic
  89. 2:46proposed the model context protocol, the
  90. 2:48MCP, which would standardize the ways
  91. 2:50that we interact with all of these
  92. 2:52external softwares, everybody was like,
  93. 2:54amazing. Wow, so good. Now your agent is
  94. 2:57able to have access to all of these
  95. 2:59different external softwares and tools
  96. 3:01using the same protocol. Within just
  97. 3:03months that the number of tools that you
  98. 3:04can give your AI agents now, which are
  99. 3:06called MCP servers massively exploded to
  100. 3:08thousands. Now there's over 20,000
  101. 3:10pre-built MCP servers that you can just
  102. 3:12grab and give your AI agent. Here is
  103. 3:14just a small sample. Now, anybody can
  104. 3:16easily write them, publish them, and use
  105. 3:18them. Let me actually show you exactly
  106. 3:20how easy it is to use an MCP server.
  107. 3:22Like say for example, I'm looking at
  108. 3:23these MCP servers here. You can pick any
  109. 3:25of them, but let's just pick like Alpha
  110. 3:27Vantage for example. It is a MCP server
  111. 3:29that enables LLMs and aentic workflows
  112. 3:31to seamlessly interact with realtime and
  113. 3:33historical stock market data through the
  114. 3:35model context protocol. Okay, great. So,
  115. 3:37you can get access to stock market data.
  116. 3:41You copy this stuff which allows your LM
  117. 3:43application to have access to the MCP
  118. 3:45server. copy it into your LM application
  119. 3:47like for example cloud code if you want
  120. 3:49to give the tool to cloud code and voila
  121. 3:51you'll be able to have access to the MCP
  122. 3:54server and all the different tools that
  123. 3:55are defined on the server then you can
  124. 3:57do something like plot the coffee stock
  125. 4:01market prices for the past 10 years then
  126. 4:05Claude will say Claude wants to use the
  127. 4:07coffee function from Alpha Vantage are
  128. 4:09we going to allow this and we'll be like
  129. 4:10okay allow once using that tool was able
  130. 4:13to get the coffee price data and now
  131. 4:14it's going to created interactive
  132. 4:16visualization and there you go, coffee
  133. 4:17global prices with all of this
  134. 4:19information. Oh my god, coffee is so
  135. 4:21expensive these days. So yes, this is
  136. 4:23pretty much how easy it is to use any
  137. 4:25MCP server now. And I can also very
  138. 4:27easily switch what is called the host
  139. 4:29which is cloud code in this case to
  140. 4:30something else like NA10 for example or
  141. 4:32my own custom AI app. So let's now
  142. 4:34actually dive into the foundations of
  143. 4:36MCP so you will exactly understand what
  144. 4:38just happened here and how to build your
  145. 4:40own MCP servers as well. And by the way,
  146. 4:42MCP servers are actually even more
  147. 4:43powerful than just giving your agents
  148. 4:45tools. There's a lot more that you can
  149. 4:47actually do, which is why people are
  150. 4:48able to actually build these MCP servers
  151. 4:50and then sell them.
  152. 4:53The best course that I found which
  153. 4:54covers the foundations of MCP is
  154. 4:56Anthropic's own course in collaboration
  155. 4:58with deep learning.AI called MCP, build
  156. 5:01rich context AI apps with anthropic. The
  157. 5:03course starts with explaining basically
  158. 5:05why is MCP important. So I already
  159. 5:07covered most of that. Um but I do like
  160. 5:09the diagrams that they gave which shows
  161. 5:10that see before MCP there was fragmented
  162. 5:13AI development with each AI app you
  163. 5:16would need to have a custom
  164. 5:17implementation custom prompt logic
  165. 5:19custom tool calls and custom data
  166. 5:21access. And you got to do that for all
  167. 5:22of them versus with MCP which is a
  168. 5:24standardized AI development for each MCP
  169. 5:27compatible application you're able to
  170. 5:29have access to a lot of different types
  171. 5:31of MCP servers. For example, you can
  172. 5:33have access to a data store MCP server
  173. 5:35so that you can have access to data
  174. 5:36stores. You can have a CRM MCP server so
  175. 5:38that you can have access to CRM systems
  176. 5:41and a version control MCP server so you
  177. 5:42can have version control software. All
  178. 5:44plug-andplay and the other way around as
  179. 5:46well. Every time you build an MCP
  180. 5:47server, it would be reusable by various
  181. 5:49AI applications like a Google Drive MCP
  182. 5:52server that has access to Google Drive
  183. 5:53functions that can be used by MCP
  184. 5:56compatible AI assistant, MCP compatible
  185. 5:58AI agent, and an MCP compatible desktop
  186. 6:00app or IDE too. All right. Next
  187. 6:02important concept to understand is the
  188. 6:03client server architecture. Within the
  189. 6:06components that make up MCP, there are
  190. 6:09three different components. The host,
  191. 6:11client, and server, HCS for short. And
  192. 6:14the way that they relate to each other
  193. 6:15is the host houses the MCP client, which
  194. 6:18goes through the MCP protocol in order
  195. 6:20to have access to the MCP server. Hosts
  196. 6:23are LM applications that want access to
  197. 6:25certain data or tools through MCP. As I
  198. 6:28showed earlier, the cloud desktop is an
  199. 6:29example of a host. NA10 is also an
  200. 6:31example of a host. You can also have
  201. 6:33IDs, AI agents. Really any large
  202. 6:35language model application can be a
  203. 6:36host. Now the host wants to use the MCP
  204. 6:39servers which are the lightweight
  205. 6:40programs that expose different
  206. 6:41capabilities like the servers here like
  207. 6:43alpha advantage we talked about earlier
  208. 6:45that's able to get real-time historical
  209. 6:47stock data. Uh there's also time
  210. 6:49provides time and time zone conversion
  211. 6:51capabilities. Cherry studio that's used
  212. 6:54for LM provider support like AI
  213. 6:56assistance and conversions, documented
  214. 6:57data processing. Postgress SQL to
  215. 6:59provide readonly access to Postgress SQL
  216. 7:01databases allowing schema inspection and
  217. 7:03execution of readonly queries, GitLab,
  218. 7:05Reddist, etc., etc. There is a lot. Now,
  219. 7:08in order for the host to actually have
  220. 7:11access to these MCP servers, they need
  221. 7:13to maintain what is called an MCP
  222. 7:15client, which lives inside a host and
  223. 7:17evokes the MCP protocol to maintain that
  224. 7:20one-on-one connection with the server.
  225. 7:22Got it? MCP client lives in MCP host to
  226. 7:25invoke the MCP protocol in order to have
  227. 7:28access to the MCP servers which the
  228. 7:30players being HCS host client and
  229. 7:33server. All right, time for a little
  230. 7:34quiz for the example that I showed
  231. 7:36earlier which allowed cloud desktop able
  232. 7:38to access historical stock data alpha
  233. 7:41vantage and cloud code. Which is the
  234. 7:43host and which is the server? Put your
  235. 7:45answer into the comments. Okay, let's
  236. 7:48now talk about what are the things that
  237. 7:49are actually contained within an MCP
  238. 7:51server. Well, there are three major
  239. 7:54things. These are tools, resources, and
  240. 7:56prompt templates. TRP tools, which are
  241. 7:59the things that we've kind of seen the
  242. 8:02most of are functions and tools that can
  243. 8:04be invoked by the client. So, these are
  244. 8:06these things over here like time series,
  245. 8:09intraday, time series, daily, blah blah
  246. 8:11blah, symbol search, global quote. Also,
  247. 8:13stuff like sending a message in Gmail,
  248. 8:15calculator, retrieval and search,
  249. 8:17sending a message, updating database
  250. 8:18records. All of these are tools. But
  251. 8:21that is actually not all. You can also
  252. 8:23have what are called resources which are
  253. 8:25readonly data that is exposed by the
  254. 8:27server. So this is the data that the
  255. 8:29client is able to query but cannot
  256. 8:31change. For example, there are certain
  257. 8:32files like markdown notes that's
  258. 8:34contained within the server that you
  259. 8:36might want to read. Maybe it's tracking
  260. 8:38logs. Say for example, every time you're
  261. 8:39calling a weather app tool and you're
  262. 8:41getting that data, you might want to
  263. 8:43actually save a record of all the
  264. 8:45different weathers that you've
  265. 8:46collected. So at some point you might
  266. 8:47want to compile that together and make
  267. 8:49like a visualization or something,
  268. 8:50right? It would not be very efficient if
  269. 8:51every single time you had to go collect
  270. 8:53a weather data point. You would have to
  271. 8:55like use the tool and spam the tool over
  272. 8:57and over and over again. It could also
  273. 8:58contain just database records of stuff.
  274. 9:00Like for example, there's just like
  275. 9:02information that you're storing like
  276. 9:03contracts, meeting recordings, and
  277. 9:05notes. Then there's also prompt
  278. 9:07templates. These are structured prompt
  279. 9:09blueprints so that you don't have to
  280. 9:10force your user to have to come up with
  281. 9:12their own prompts. Like for example, if
  282. 9:14you are someone who's trying to use an
  283. 9:15MCP server that uh is able to summarize
  284. 9:18like a certain transcript from a meeting
  285. 9:20and then generate a report based upon
  286. 9:22it, you could just write something like,
  287. 9:23oh, summarize this transcript into key
  288. 9:25action items, then generate a report,
  289. 9:28right? Like you could write something
  290. 9:29like that, but that probably won't yield
  291. 9:31the best results. So instead, the people
  292. 9:32that actually built the MCP server can
  293. 9:35put within their MCP server a prompt
  294. 9:37template that is much more flushed out
  295. 9:39that's able to actually extract the
  296. 9:41insights and generate the report in a
  297. 9:42much better fashion so that the user all
  298. 9:44they have to do is input like some
  299. 9:46specifics into that template. This takes
  300. 9:48the burden of the prompt engineering
  301. 9:50from the user. As a full example, you
  302. 9:52can have like a SQLite MCP server which
  303. 9:55gives access to like a database which is
  304. 9:57the SQLite database. Don't worry if you
  305. 9:59don't know what that is. is basically
  306. 10:00just like a database that can store
  307. 10:01information. Anyways, so you can have
  308. 10:03tools that allow you to read the SQL
  309. 10:06database to insert information into the
  310. 10:08database to change the information in
  311. 10:10the database and to delete information
  312. 10:11from the database. In addition to that,
  313. 10:13there is a resource that is logging all
  314. 10:16of the changes that are happening to the
  315. 10:17database so you have a historical track
  316. 10:19record of what's happened. And there's
  317. 10:20also some prompt templates that
  318. 10:22incorporates the best practices to
  319. 10:24interact with that database. Make sense?
  320. 10:25All right, time for our next little
  321. 10:27assessment. Please answer the questions
  322. 10:28on screen now and put it into the
  323. 10:31comments so you are following along. MCP
  324. 10:33is all about giving your agents more
  325. 10:35context so that they can do really
  326. 10:36powerful things. So when you're ready to
  327. 10:38deploy your AI agents with a lot more
  328. 10:40functionalities and you want really good
  329. 10:41UI, NA10 plus Bolt is a really powerful
  330. 10:45combination and you don't need to write
  331. 10:46a single line of code or you can just
  332. 10:48directly build via Bolt. Bolt can take
  333. 10:50you from an idea to a fully functional
  334. 10:52product within minutes. Whether that is
  335. 10:54software for your team, an internal tool
  336. 10:56or a website. It has everything that you
  337. 10:58need, whether that's hosting, backend,
  338. 11:00database, off, domains, and even SEO.
  339. 11:03One of the things I really like about
  340. 11:04Bald is that it seamlessly handles the
  341. 11:06entire process so that you can focus on
  342. 11:08building real things as opposed to
  343. 11:11handling infrastructure. You don't need
  344. 11:12to handle like setup, configs, war,
  345. 11:14external tools. Bolt puts the most
  346. 11:16powerful tools used by developers into
  347. 11:18your hands with an intuitive chat
  348. 11:20interface. It really is as simple as
  349. 11:22this. So, here we have the Bolt new
  350. 11:24landing page. And we can just type
  351. 11:26something like build an app that blurs
  352. 11:29out people's faces in pictures and just
  353. 11:33click build now and then let it do its
  354. 11:36thing. Of course, you can go into a lot
  355. 11:38more details and have a lot of like
  356. 11:39specific specs within the app as well.
  357. 11:42And look at that. Here we have this
  358. 11:44application and let's actually try it
  359. 11:45out. We will put an image. Here is a
  360. 11:49picture of me when I was still a little
  361. 11:50Tina. And there you go. We have the
  362. 11:52original and then we have the blurred
  363. 11:54version. And you can download the
  364. 11:56blurred image as well. As well as you
  365. 11:58keep building, you can include a lot
  366. 11:59more features and tweak things however
  367. 12:01it is that you want. Let Bald do the
  368. 12:03heavy lifting so you can focus on your
  369. 12:04vision instead of fighting bugs. You can
  370. 12:06try out Bald V2 for free today at this
  371. 12:09link over here, also linked in
  372. 12:10description. Now, back to the video.
  373. 12:12Okay, so the final concept that I want
  374. 12:14to cover for MCP before diving into
  375. 12:16actually building them is the
  376. 12:18communication life cycle. The way that
  377. 12:19the MCP client and the MCP server
  378. 12:22communicate with each other. The first
  379. 12:23step is the initialization when the MCP
  380. 12:25client interacts with the MCP server
  381. 12:27being like, hey, I would like things
  382. 12:29from you now and they initiate a
  383. 12:31connection. Then you have the message
  384. 12:33exchange phase when the MCP client makes
  385. 12:35a request to the server. for example
  386. 12:37like I want to use this tool and the MCP
  387. 12:40server goes like okay and gives it back
  388. 12:42which is the response. So you do that
  389. 12:44exchange the messages and then finally
  390. 12:46there's the termination phase when the
  391. 12:48client and the server terminate and stop
  392. 12:51interacting with each other. Now the way
  393. 12:52that the messages are done is called a
  394. 12:54transport. The transport handles the
  395. 12:56underlying mechanics of how messages are
  396. 12:58sent and received between the client and
  397. 13:00the server. And there are two different
  398. 13:01categories of how transports work. First
  399. 13:03one is when the server is running
  400. 13:05locally like your MCP server and your
  401. 13:07host all of this is within a single
  402. 13:09machine. The analogy that I like to give
  403. 13:11here is say that your MCP client is like
  404. 13:13the diner and the chef is like your MCP
  405. 13:16server who is going to be the one
  406. 13:18preparing the food for your diner. Now
  407. 13:19for the server and the client running
  408. 13:21locally, it'll be like just you and your
  409. 13:23friend cooking at home with like a
  410. 13:24notebook. You as the chef is just like
  411. 13:26cooking it and maybe you're just like
  412. 13:28writing down what it is that you're
  413. 13:29doing and like just passing a note to
  414. 13:30your friend the diner for some reason.
  415. 13:32So they're like keeping updated about
  416. 13:34what's happening while you're cooking.
  417. 13:35That is like the standard out a running
  418. 13:37log that's just like living locally
  419. 13:39between you guys. You don't need to rely
  420. 13:41on like a waiter or a cashier or
  421. 13:43anything like that. The other category
  422. 13:44of transports is for remote servers. So
  423. 13:46these are servers that live in other
  424. 13:48machines on the cloud somewhere else.
  425. 13:50And there are two different approaches
  426. 13:51to interact with a remote server. The
  427. 13:53first one is called HTTP plus SSE which
  428. 13:55is server sent events. Don't worry so
  429. 13:57much about the terminologies. It
  430. 13:58supports what is called a stateful
  431. 14:00connection. Going back to that analogy,
  432. 14:02in this case, you would have to like
  433. 14:03leave your house and say you're at a
  434. 14:04restaurant. You, as a diner, would sit
  435. 14:06down at the restaurant. You're the
  436. 14:08client, and you would interact with a
  437. 14:10waiter, which is a server, and you would
  438. 14:12tell it stuff like, "Oh, here's my
  439. 14:13table. Here's the appetizers that I want
  440. 14:15to order, blah, blah, blah." And then
  441. 14:16your server would like bring you those
  442. 14:17things, and say if you want to order
  443. 14:19like the same drink again, you could
  444. 14:20just say something like, "I'll like to
  445. 14:21have the same drink again." And you
  446. 14:22don't have to like exactly specify
  447. 14:24whatever it is that you ordered the
  448. 14:25first time. So this is called a stateful
  449. 14:27connection because the server maintains
  450. 14:28the state of it like remembering the
  451. 14:30knowledge the memory of what you ordered
  452. 14:31all this information throughout the
  453. 14:33entire interaction. Now there's also
  454. 14:35what is called stateless stateful
  455. 14:37stateless. This kind of transport
  456. 14:38supported by a streamable HTTP. To be
  457. 14:41stateless is sort of like if you go to a
  458. 14:43fast food restaurant as opposed to like
  459. 14:45a sitdown restaurant. Like if you're at
  460. 14:47McDonald's for example, you would order
  461. 14:48like a Big Mac and I don't know like
  462. 14:50fries. But then like if you want to get
  463. 14:52something else, you want to also get
  464. 14:53like chicken nuggets, you would actually
  465. 14:55have to go and like input your order all
  466. 14:57over again and get like a new ticket and
  467. 14:59you have to wait again. The cashier
  468. 15:00that's clashing you out, the server
  469. 15:01doesn't actually remember everything
  470. 15:03that you ordered previously. So that's
  471. 15:05why it's called stateless because it
  472. 15:06doesn't remember all of the knowledge
  473. 15:08that came with your prior interactions.
  474. 15:10Each request is independent of each
  475. 15:12other. So without going into way too
  476. 15:13much detail about this like the
  477. 15:14streamable HTTP is the preferred
  478. 15:17transport method because it's able to
  479. 15:19support both stateful and stateless and
  480. 15:21there are instances in which you want to
  481. 15:23have like stateful interactions and
  482. 15:25instances in which you want to have
  483. 15:26stateless interactions as well. So this
  484. 15:28section might sound a little bit like
  485. 15:29detailed and theoretical but it does
  486. 15:31become very important when we're
  487. 15:33actually going to be using our MCP
  488. 15:35servers and building our MCP servers
  489. 15:37which is right after I give you this
  490. 15:38final little assessment.
  491. 15:44Okay, I'm going to show you the no code
  492. 15:46implementation of building your very own
  493. 15:47MCP server now. And we're going to be
  494. 15:49doing this by using n10. So you can do
  495. 15:52stuff like send an email to Tina with
  496. 15:55the address tina@ lonelyotopus.com
  497. 15:59asking her if she wants to grab a coffee
  498. 16:04sometime.
  499. 16:07Execute workflow. Press enter. goes
  500. 16:11through this. We can see that it's
  501. 16:12accessing the AI agent is accessing the
  502. 16:14MCP client which is here. It contains
  503. 16:17the ability to have a calculator and
  504. 16:19also the tool that allows to send a
  505. 16:21message to Gmail. So, it's able to
  506. 16:23access the tool to send a message to
  507. 16:25Gmail so that you will be able to get
  508. 16:27this email. Hi Tina, I hope this message
  509. 16:30finds you well. I was wondering if you
  510. 16:31like to grab a coffee sometime. Let me
  511. 16:32know what works for you. Best regards.
  512. 16:33Email was automatically sent with NATO.
  513. 16:35Okay. So, how to build this from
  514. 16:37scratch? start off with a fresh new
  515. 16:39workflow using NA10. If you have no idea
  516. 16:41what NA10 is and you're not familiar
  517. 16:43with it at all, check out this video
  518. 16:44over here. I go into details about how
  519. 16:46to build AI agents using NA10. I'm not
  520. 16:48going to go into this here, so the video
  521. 16:49is not super long, but basically it's a
  522. 16:51no code low code tool that allows you to
  523. 16:53build workflows. Also allows you to
  524. 16:54build MCP servers. So you have your
  525. 16:56server trigger over here and you want to
  526. 16:58add some tools. Like for example, we can
  527. 17:00add the calculator tool, which is just
  528. 17:02like the ability to calculate things and
  529. 17:04do math. Not super innovative but yes.
  530. 17:07And then we can also add the email tool.
  531. 17:09So maybe like a Gmail tool. And in here
  532. 17:13we need to set up the credentials like
  533. 17:14create new credential. You can kind of
  534. 17:16go through the process. So all you have
  535. 17:19to do is like sign into your Gmail um
  536. 17:21which I will do here. And then you'll
  537. 17:23have the account that is connected. Yay.
  538. 17:26Then um because we just want to send
  539. 17:28emails here. So the operation here is
  540. 17:29going to be send. So for two we're going
  541. 17:31to let the model decide these. So, let
  542. 17:33the model decide the subject. Let's al
  543. 17:36also let the model decide uh we'll have
  544. 17:38the email type HTML and the message.
  545. 17:42Let's also let the model decide that as
  546. 17:43well. Wonderful. Amazing. So simple.
  547. 17:45Right now, let's save this. Now, in
  548. 17:48order to actually use our server, we
  549. 17:50want to come here, make sure that it's
  550. 17:52active. So, the server itself is going
  551. 17:54to be active. And we want to take this
  552. 17:57production URL. So, this allows access
  553. 17:59to the server. Now I have another
  554. 18:01workflow over here which is an AI agent
  555. 18:03in which you can prompt it directly and
  556. 18:05respond also attach the open AI chat
  557. 18:07model here. So in order to add the MCP I
  558. 18:10can click the MCP client tool the
  559. 18:12endpoint we're going to copy paste what
  560. 18:14we got from the MCP server and here in
  561. 18:17the server transport see we have the
  562. 18:19server send events which is deprecated
  563. 18:20and the HTTP streamable the two
  564. 18:22different transport right so we're going
  565. 18:24to click HTTP streamable transport and
  566. 18:26we don't need to change anything else
  567. 18:28over here. Wow. Now we have the MCP
  568. 18:30client. So let's actually test it out.
  569. 18:32Let's actually do math this time.
  570. 18:34Calculate 10
  571. 18:3711 * 99.
  572. 18:41We can see here that the AI agent goes
  573. 18:44over here. Human says calculate 11 * 99.
  574. 18:46It goes to the MCP client and it uses
  575. 18:48the tool cal the calculator tool which
  576. 18:50we have. And it says useful for getting
  577. 18:52the results of a math expression. And
  578. 18:54then the open AI chat model was able to
  579. 18:55get the results by using the MCP client.
  580. 18:58And then we were able to get the final
  581. 18:59result here with the output being the
  582. 19:01result of 11 multipi by 99 is 1,089.
  583. 19:05Yay. See how easy it is. There's a lot
  584. 19:08that you can do with this. Of course,
  585. 19:09you can add like so many more tools into
  586. 19:11your MCP server. And you also don't need
  587. 19:13to have NAN as your host. We can easily
  588. 19:16switch out the host for something else.
  589. 19:17Like for example, maybe you want to
  590. 19:19change out the host for cloud desktop.
  591. 19:20So what we need to do is go on cloud
  592. 19:23desktop. So go on help, type in
  593. 19:25developer. So switch to developer mode
  594. 19:28and then under developer open the claw
  595. 19:30desktop config file. So this should be
  596. 19:33empty. You can open it with whatever
  597. 19:35text editor. So I'm here I'm just using
  598. 19:36VS code for it. And you want to copy
  599. 19:38paste this. So this is defined MCP
  600. 19:40server and we're using NAN because
  601. 19:42that's where the server is and we are
  602. 19:44using SSE over here. And over here just
  603. 19:48paste the endpoint here. So command S or
  604. 19:52control S. Save it. Restart Claude and
  605. 19:56wow amazing. We have NA10 over here
  606. 19:58under the tools and includes send a
  607. 20:00message in Gmail and the calculator. And
  608. 20:02then you can do the same thing here.
  609. 20:04Send an email to Tina@ lonely
  610. 20:08octopus.com
  611. 20:09asking her to host an octopus party.
  612. 20:13Click enter.
  613. 20:15See it'll say over here Claude wants to
  614. 20:17use send a message in Gmail from NA10.
  615. 20:19And you can click allow once to allow it
  616. 20:21to use the tool from the MCP server. And
  617. 20:24moment of truth, we have over here, "Hi
  618. 20:27Tina, I hope this message finds you
  619. 20:28well. I want to reach out to see if
  620. 20:29you'd be interested in hosting an
  621. 20:30octopus party. I think it'll be a
  622. 20:32wonderful event and you'll be the
  623. 20:33perfect host for it. Let me know if
  624. 20:34you're available and interested. Looking
  625. 20:35forward to hearing from you. Best
  626. 20:36regards. Yes, automatically sent with
  627. 20:38NAT." Voila. See, super easy, right?
  628. 20:41Okay. So, now I want to show you the
  629. 20:42code implementation of doing this. Now
  630. 20:44with code you are able to do a lot more
  631. 20:47than this no code approach. I think the
  632. 20:49no code approach is amazing. It's great,
  633. 20:50right? It's really accessible and easy
  634. 20:52to use. Remember we talked about earlier
  635. 20:53about having like additional resources
  636. 20:55and also prompt templates. You as of
  637. 20:57right now cannot do that using NAT. You
  638. 21:00can't actually put in the prompt
  639. 21:01templates and specify those and you also
  640. 21:02can't put in additional resources as
  641. 21:04well. And in the code implementation you
  642. 21:06can. So I'm not going to go and show you
  643. 21:08every single line of code right now
  644. 21:10because I don't want to make this video
  645. 21:11like super duper long. But I want to
  646. 21:13show you like the snippet of these are
  647. 21:14the tools that we're going to be using
  648. 21:16and this is how you can actually get it
  649. 21:17to run. Okay. So for this code demo,
  650. 21:19we're going to be using cloud desktop as
  651. 21:22the host and I'm going to showcase a
  652. 21:24simple MCP server that's able to provide
  653. 21:26tools, resources, and prompts for
  654. 21:28working with Google Sheets and Google
  655. 21:30forms. So we can tell that the MC3
  656. 21:32server is connected because if you go to
  657. 21:34G Sheets over here, which is the MCP
  658. 21:36server name, you can see all the tools
  659. 21:37that are contained over here. So I have
  660. 21:39a lot of different spreadsheets that are
  661. 21:41available. So let's actually ask list
  662. 21:42out the spreadsheets that are available.
  663. 21:49It will want to use the list
  664. 21:50spreadsheets. So this is the tool and it
  665. 21:53tells us that we have 20 spreadsheets
  666. 21:55that are available in the Google account
  667. 21:57like NA10 form, anonymous feedback, the
  668. 22:00only octis alumni blah blah blah etc.
  669. 22:02All of these different ones. Would you
  670. 22:04like me to read the data from any of
  671. 22:05these spreadsheets here? I can say use
  672. 22:08read sheet to get the columns
  673. 22:13of the anonymous
  674. 22:17feedback spreadsheet.
  675. 22:20So it wants to use read sheet and this
  676. 22:22is actually a resource. So it allows you
  677. 22:26it's a readonly file where you can
  678. 22:28actually read what the column names are.
  679. 22:30You can see the spreadsheet ID and the
  680. 22:33information that's contained over there.
  681. 22:35And finally, I want to show you what it
  682. 22:36looks like to use some of the prompt
  683. 22:38templates. So, if we click over here,
  684. 22:40add from sheets, here are some of the
  685. 22:43prompt templates. So, let's click
  686. 22:44analyze sheet data. And we can say use
  687. 22:46this to analyze the anonymous
  688. 22:52pretty spell anonymous
  689. 22:55feedback
  690. 22:57spreadsheet
  691. 22:59and create a dashboard just to make it
  692. 23:02look pretty.
  693. 23:04And we can see here is what the prompt
  694. 23:06template actually looks like. And it
  695. 23:09shall do this. Let it do its thing. So
  696. 23:12while it's doing that, let me actually
  697. 23:13show you what the code looks like and
  698. 23:14how to run this MCP server. Here's the
  699. 23:16readme. If you follow these
  700. 23:17instructions, you'll be able to run this
  701. 23:18MCP server yourself. You can see that
  702. 23:21the tools that it contains include
  703. 23:23things like list spreadsheet, read the
  704. 23:25sheet, write the sheet, append the
  705. 23:26sheet, etc. It also has the resource
  706. 23:29where you can use the sheet in order to
  707. 23:31access sheet data as the resource. And
  708. 23:33it has prompt templates including
  709. 23:34analyze sheet data which is template for
  710. 23:36analyzing sheet data. Creating report
  711. 23:38template. So a template for creating
  712. 23:40report spreadsheets and form to sheet
  713. 23:42which is a template for form to sheet
  714. 23:44workflow in the code itself. Not going
  715. 23:47into too much detail over here. You can
  716. 23:50see that after we define the credentials
  717. 23:52the way that you can define these tools
  718. 23:54that you can decorate the function. So
  719. 23:56you know that's MCP tool. So this one is
  720. 23:58the function for listing the
  721. 23:59spreadsheet. This one is for reading the
  722. 24:01sheet etc. And then the way that you can
  723. 24:04show that it's a resource is that you
  724. 24:06just decorate it so that you know that
  725. 24:07it's a resource. So get sheet resource.
  726. 24:10And here are the prompt templates
  727. 24:12themselves. So they're actually
  728. 24:13hardcoded directly in these different
  729. 24:15prompt templates. So let's go back to
  730. 24:17claude here and see what is happening
  731. 24:19there. Great. We see that it analyzed
  732. 24:22the anonymous feedback spreadsheet. So I
  733. 24:25read that information, analyzed the data
  734. 24:27using the template that we provided and
  735. 24:29it actually created a dashboard as well.
  736. 24:31So here is the dashboard. So total
  737. 24:34response is 43. The average MPS score
  738. 24:37which is the net promoter score is 4.9
  739. 24:395.98 blah blah blah all these things. So
  740. 24:42it tells you information about this
  741. 24:45different categories uh how the study
  742. 24:47plan progress tracker is qualitative
  743. 24:50information about what attracted
  744. 24:52students to our boot camp and insights
  745. 24:55as well. So different types of
  746. 24:57information including what works well
  747. 24:59and what doesn't work well etc etc. So
  748. 25:01yeah there you go. This is the code demo
  749. 25:04of an MCP server. I hope you can see
  750. 25:06that there's so much potential for so
  751. 25:08many different types of things that you
  752. 25:09can build now that you have the ability
  753. 25:11to create your own MCP servers. You can
  754. 25:13of course also change your host. You use
  755. 25:14this MCP server with NA10 uh with your
  756. 25:18own agent with your any large language
  757. 25:20model application that you build as
  758. 25:22well. The possibilities are endless. So
  759. 25:24if you do want to dig deeper into this,
  760. 25:26I highly recommend that you actually
  761. 25:27check out the anthropic deep learning AI
  762. 25:29course that goes into the exact
  763. 25:31implementation step by step how to build
  764. 25:33this entire MCP server. It is really,
  765. 25:35really, really powerful. There are of
  766. 25:36course a lot of other ways that you can
  767. 25:37implement MCP servers as well, both code
  768. 25:40and no code approaches. I also really
  769. 25:41like this one by Network Chuck where he
  770. 25:44goes into a lot of detail about how to
  771. 25:45use Docker in order to do this. So I'm
  772. 25:47going to link his video below as well. I
  773. 25:48really recommend that you check it out
  774. 25:49if you're interested in implementing it
  775. 25:51using Docker. But for this video, I'm
  776. 25:53going to end it here. I hope that you
  777. 25:54now have an excellent fundamental
  778. 25:56understanding of MCP and how to actually
  779. 25:58build them yourself, too. Here is the
  780. 26:00final little assessment. Please answer
  781. 26:01these questions and put them into the
  782. 26:03comments.
  783. 26:06Hope you enjoyed this video and thank
  784. 26:08you so much for staying until the end.
  785. 26:09I'll see you guys in the next video for
  786. 26:11live stream.
  787. 26:17[Music]

About this transcript

This page contains the full transcript of MCP In 26 Minutes (Model Context Protocol) by Tina Huang, generated from the public captions YouTube serves with the video. The transcript has 5,477 words across 787 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.