WEBVTT 00:00:00.000 --> 00:00:02.500 Welcome to the deep dive. Today, we're tackling 00:00:02.500 --> 00:00:04.360 something really exciting in the world of AI, 00:00:04.780 --> 00:00:06.919 how it's moving from just thinking to actually 00:00:06.919 --> 00:00:10.300 doing. You know, AI, especially the big models 00:00:10.300 --> 00:00:12.679 like ChatGPT and Claude, they've become incredibly 00:00:12.679 --> 00:00:14.820 good thinkers, right? They can draft that perfect 00:00:14.820 --> 00:00:18.280 email, map out a whole business plan, even spot 00:00:18.280 --> 00:00:20.199 bugs in code. It's like having a brilliant consultant 00:00:20.199 --> 00:00:23.980 on tap. But here's the snag, the big disconnect, 00:00:24.260 --> 00:00:26.260 what some people call the last mile problem, 00:00:26.640 --> 00:00:29.379 that AI. It writes the email, but it can't hit 00:00:29.379 --> 00:00:32.149 send. It analyzes sales data, sure, but it can't 00:00:32.149 --> 00:00:34.329 actually go into your CRM and pull it. It's a 00:00:34.329 --> 00:00:36.929 thinker, a planner, but it's stuck in a kind 00:00:36.929 --> 00:00:39.009 of digital sandbox. It can't really interact 00:00:39.009 --> 00:00:40.990 with all the tools we use every day. Exactly. 00:00:41.259 --> 00:00:43.920 That limitation has historically kept AI, even 00:00:43.920 --> 00:00:45.920 the really advanced stuff, in the role of an 00:00:45.920 --> 00:00:48.119 advisor, a very smart advisor. But still, just 00:00:48.119 --> 00:00:49.979 advising there were early attempts to bridge 00:00:49.979 --> 00:00:51.780 this gap. People started connecting these language 00:00:51.780 --> 00:00:54.719 models to external tools using APIs, application 00:00:54.719 --> 00:00:56.560 programming interfaces. That was a key step, 00:00:56.659 --> 00:00:59.320 letting AI do things like search the web. But 00:00:59.320 --> 00:01:01.600 trying to scale that, it quickly ran into some 00:01:01.600 --> 00:01:04.079 serious, almost crippling, complexity. Right. 00:01:04.079 --> 00:01:05.739 And that's where today's topic comes in and why 00:01:05.739 --> 00:01:07.739 it's so interesting. We're doing a deep dive 00:01:07.739 --> 00:01:10.700 into a new standard that aims to fix this really 00:01:10.569 --> 00:01:13.370 elegantly and at scale. It's called the Model 00:01:13.370 --> 00:01:16.989 Context Protocol or MCP. Our mission today is 00:01:16.989 --> 00:01:19.409 to figure out how MCP represents this huge leap, 00:01:19.790 --> 00:01:23.090 moving us from just talking to AI to having AI 00:01:23.090 --> 00:01:27.069 actually automate complex tasks for us. And crucially, 00:01:27.409 --> 00:01:30.269 what does this mean for you? So to really get 00:01:30.269 --> 00:01:33.069 why MCP is such a big deal, maybe we should start 00:01:33.069 --> 00:01:35.329 with the problem it solves. What was that first 00:01:35.329 --> 00:01:37.989 idea for getting LLMs to use tools? You mentioned 00:01:37.989 --> 00:01:40.450 APIs. Yeah, the initial thought seemed pretty 00:01:40.450 --> 00:01:42.650 straightforward. If the AI needs to, say, check 00:01:42.650 --> 00:01:44.790 your calendar or look up a customer in Salesforce, 00:01:45.090 --> 00:01:46.769 just hook it up to the Google Calendar API or 00:01:46.769 --> 00:01:49.609 the Salesforce API. Simple, right? Well, no, 00:01:49.829 --> 00:01:51.590 that idea just completely shattered when people 00:01:51.590 --> 00:01:53.870 tried to do it at scale. The fundamental issue 00:01:53.870 --> 00:01:56.469 is there's no universal language for APIs out 00:01:56.469 --> 00:01:58.609 there in the digital world. Every single service, 00:01:58.689 --> 00:02:00.810 every app speaks its own lightly different dialect. 00:02:01.030 --> 00:02:02.890 OK, so it's not like USB where everything just 00:02:02.890 --> 00:02:04.890 plugs in? Not at all. It's more like needing 00:02:04.890 --> 00:02:07.569 a unique adapter and maybe even a voltage converter 00:02:07.569 --> 00:02:10.360 for every single device you own. This lack of 00:02:10.360 --> 00:02:12.860 standardization created a huge mess. You had 00:02:12.860 --> 00:02:16.680 things like inconsistent authentication. One 00:02:16.680 --> 00:02:20.680 tool needs an API key. Another uses 002 .0, which 00:02:20.680 --> 00:02:23.560 is that whole multi -step secure login dance. 00:02:23.680 --> 00:02:26.360 A third wants a bear token. So developers had 00:02:26.360 --> 00:02:29.819 to build custom logic for every single authentication 00:02:29.819 --> 00:02:32.560 method. Then there were divergent data structures. 00:02:32.879 --> 00:02:35.259 What HubSpot calls a user and the information 00:02:35.259 --> 00:02:37.039 it includes looks totally different from what 00:02:37.039 --> 00:02:39.599 Zendesk calls a user. So you're manually mapping 00:02:39.599 --> 00:02:42.120 fields, translating data back and forth. It's 00:02:42.120 --> 00:02:44.680 tedious. And don't forget varying error handling 00:02:44.680 --> 00:02:47.259 and rate limits. how one API tells you something 00:02:47.259 --> 00:02:49.120 went wrong, or how often you're allowed to call 00:02:49.120 --> 00:02:52.560 it, is unique. More custom logic for every single 00:02:52.560 --> 00:02:55.020 endpoint. That sounds brittle. Extremely. It 00:02:55.020 --> 00:02:57.259 became a maintenance nightmare. Imagine Notion 00:02:57.259 --> 00:02:59.039 updates its API, which happens all the time. 00:02:59.099 --> 00:03:01.240 Poof! Your custom integration breaks. Multiply 00:03:01.240 --> 00:03:04.039 that by dozens, maybe hundreds of tools. You 00:03:04.039 --> 00:03:06.060 end up with a system that's incredibly fragile, 00:03:06.300 --> 00:03:08.180 super expensive to maintain, and just doesn't 00:03:08.180 --> 00:03:11.300 scale. So it really is like that brilliant CEO 00:03:11.300 --> 00:03:14.180 needing a personal translator for every single 00:03:14.180 --> 00:03:16.860 department head. Massively inefficient. I mean, 00:03:17.039 --> 00:03:20.699 I've felt that pain just trying to get two regular 00:03:20.699 --> 00:03:23.639 software tools to sync properly. It's clear something 00:03:23.639 --> 00:03:26.460 had to change. A universal standard was desperately 00:03:26.460 --> 00:03:28.740 needed. Absolutely. The friction was just becoming 00:03:28.740 --> 00:03:31.580 unbearable. Which brings us neatly to the model 00:03:31.580 --> 00:03:34.879 context protocol, MCP. This is positioned as 00:03:34.879 --> 00:03:37.139 the universal standard, right? A real leap forward. 00:03:37.379 --> 00:03:40.280 It absolutely is. Anthropic introduced MCP in 00:03:40.280 --> 00:03:43.280 late 2024, and crucially, they released it as 00:03:43.280 --> 00:03:46.139 an open standard. Its whole purpose is to provide 00:03:46.139 --> 00:03:49.360 one single standardized way for AI models to, 00:03:49.800 --> 00:03:52.240 first, discover what tools are available, second, 00:03:52.400 --> 00:03:54.680 understand how to use them, and third, actually 00:03:54.680 --> 00:03:56.719 use them. OK, so like a universal translator. 00:03:56.780 --> 00:03:59.240 That's a good analogy. Or think of it like a 00:03:59.240 --> 00:04:01.919 universal power adapter for AI. Instead of needing 00:04:01.919 --> 00:04:04.699 a different plug for every or every API, you've 00:04:04.699 --> 00:04:08.000 got one adapter MCP. The AI only needs to learn 00:04:08.000 --> 00:04:10.240 to speak this one language, the MCP language. 00:04:10.620 --> 00:04:13.099 And any tool or service like Notion or Asana 00:04:13.099 --> 00:04:15.460 or whatever, if it wants to be used by the AI, 00:04:15.560 --> 00:04:18.300 it needs to implement an MCP -compatible server. 00:04:18.589 --> 00:04:20.649 Ah, so the responsibility shifts. Completely, 00:04:20.769 --> 00:04:22.769 it flips the script. Okay, let's maybe unpack 00:04:22.769 --> 00:04:25.329 this a bit more. You mentioned clients and servers. 00:04:25.709 --> 00:04:29.350 What are the three core components making this 00:04:29.350 --> 00:04:31.730 MCP ecosystem actually work? Right, there are 00:04:31.730 --> 00:04:34.509 basically three key pieces. First, the MCP client. 00:04:34.649 --> 00:04:36.949 That's the AI application itself, chat -chip 00:04:36.949 --> 00:04:39.110 -a -tee, Claude, maybe a custom agent someone 00:04:39.110 --> 00:04:41.709 builds. It figures out what you want, decides 00:04:41.709 --> 00:04:43.949 if a tool is needed, and then makes the request 00:04:43.949 --> 00:04:46.389 using the MCP protocol. Second, you have the 00:04:46.389 --> 00:04:48.550 MCP server. This is the part built and run by 00:04:48.550 --> 00:04:51.230 the tool provider, Notion, Asana, maybe your 00:04:51.230 --> 00:04:53.750 company's internal database team. Its job is 00:04:53.750 --> 00:04:56.089 to advertise what capabilities it offers, the 00:04:56.089 --> 00:04:58.350 tools, the data, the actions the AI is allowed 00:04:58.350 --> 00:05:00.189 to use, and then actually do those things when 00:05:00.189 --> 00:05:03.310 the client asks. And the third piece is the MCP 00:05:03.310 --> 00:05:06.110 protocol itself. That's the rule book. The standardized 00:05:06.110 --> 00:05:08.149 set of messages for how the client and server 00:05:08.149 --> 00:05:10.389 talk to each other, it defines how the client 00:05:10.389 --> 00:05:12.850 asks, hey, what can you do? That's technically 00:05:12.850 --> 00:05:15.689 a list tool's request. And now it says, OK, do 00:05:15.689 --> 00:05:17.850 this specific thing, which is a call tool request. 00:05:17.990 --> 00:05:20.689 Got it. Client asks, server answers and acts. 00:05:20.910 --> 00:05:23.439 Protocol is the language they both speak. Exactly. 00:05:23.800 --> 00:05:25.779 And this leads to that crucial paradigm shift 00:05:25.779 --> 00:05:29.040 we mentioned. The burden of creating and maintaining 00:05:29.040 --> 00:05:32.000 the translator, the complex API integration logic 00:05:32.000 --> 00:05:34.819 moves away from the AI developer and onto the 00:05:34.819 --> 00:05:38.120 tool provider. So if Asana wants AI agents to 00:05:38.120 --> 00:05:41.379 use its platform, Asana is now motivated to build 00:05:41.379 --> 00:05:44.459 a really good, reliable MCP server. It creates 00:05:44.459 --> 00:05:47.000 this healthy competition. Service providers want 00:05:47.000 --> 00:05:49.360 their tools to be easily accessible and powerful 00:05:49.360 --> 00:05:51.560 for AI because that's the future. They handle 00:05:51.560 --> 00:05:54.189 their own API complexities behind their MCP server. 00:05:54.269 --> 00:05:56.490 That makes so much sense. It distributes the 00:05:56.490 --> 00:05:58.910 workload and aligns incentives. Okay, so we've 00:05:58.910 --> 00:06:01.829 seen the chaos of direct APIs and we've met MCP, 00:06:01.949 --> 00:06:04.189 the potential solution. What does this actually 00:06:04.189 --> 00:06:06.949 mean in practice? What are the big game -changing 00:06:06.949 --> 00:06:09.829 advantages here? The advantages are pretty profound, 00:06:10.129 --> 00:06:12.250 actually. First off, you get true model agnosticism. 00:06:12.509 --> 00:06:14.389 What that means is your tool integrations aren't 00:06:14.389 --> 00:06:17.350 locked into one specific AI model anymore. You 00:06:17.350 --> 00:06:19.370 can swap out Claude for Gemini or maybe some 00:06:19.370 --> 00:06:21.509 new open source model that comes along next year 00:06:21.509 --> 00:06:24.310 with minimal fuss. You use the best AI for the 00:06:24.310 --> 00:06:26.910 job, and you avoid getting locked into one vendor's 00:06:26.910 --> 00:06:29.110 ecosystem. That's huge for businesses. Think 00:06:29.110 --> 00:06:31.009 about future -proofing. You're not betting the 00:06:31.009 --> 00:06:34.069 farm on open AI or Anthropic if a better, cheaper, 00:06:34.189 --> 00:06:36.370 or more specialized model appears. You just plug 00:06:36.370 --> 00:06:39.730 it in, assuming it's MCP. Precisely. That flexibility 00:06:39.730 --> 00:06:42.370 is a massive strategic advantage. Second big 00:06:42.370 --> 00:06:46.329 win. Radical simplicity and scalability. Instead 00:06:46.329 --> 00:06:48.870 of that tangled mess of point -to -point API 00:06:48.870 --> 00:06:51.250 connections we talked about, you get a clean 00:06:51.250 --> 00:06:53.709 hub -and -spoke model. Your AI client learns 00:06:53.709 --> 00:06:57.209 one protocol MCP and boom, it can talk to any 00:06:57.209 --> 00:06:59.970 tool that also speaks MCP. Adding a new tool 00:06:59.970 --> 00:07:02.550 isn't some multi -week coding nightmare anymore. 00:07:02.610 --> 00:07:04.610 It's closer to just configuring it. Click, click, 00:07:04.730 --> 00:07:07.180 done. Okay. Simpler, more scalable. What else? 00:07:07.699 --> 00:07:09.480 Third, and this is really interesting for companies, 00:07:10.060 --> 00:07:13.379 unlocking internal systems. Because MCP is an 00:07:13.379 --> 00:07:15.920 open standard, businesses can build MCP servers 00:07:15.920 --> 00:07:18.800 for their own internal, maybe even ancient, proprietary 00:07:18.800 --> 00:07:20.839 systems. Wait, so you could have an AI talking 00:07:20.839 --> 00:07:23.240 to that weird old inventory database in the back 00:07:23.240 --> 00:07:25.360 office that nobody wants to touch? Potentially, 00:07:25.560 --> 00:07:29.060 yes, or interacting with a custom -built HR portal 00:07:29.060 --> 00:07:31.740 or pulling diagnostic logs from internal servers, 00:07:32.259 --> 00:07:34.480 all using the same standard interface that it 00:07:34.480 --> 00:07:37.139 uses to talk to Google Drive or Slack. But how 00:07:37.139 --> 00:07:40.360 robust is that for those really quirky legacy 00:07:40.360 --> 00:07:43.000 systems? Does the old system need a total rewrite? 00:07:43.240 --> 00:07:46.230 Great question. No, usually not. You build the 00:07:46.230 --> 00:07:48.990 MCP server as layer on top of the legacy system. 00:07:49.269 --> 00:07:51.670 Think of it like a modern facade on an old building. 00:07:52.149 --> 00:07:54.610 The MCP server takes the standard MCP requests 00:07:54.610 --> 00:07:56.610 from the AI and translates them into whatever 00:07:56.610 --> 00:07:58.689 weird calls the old system understands. Then 00:07:58.689 --> 00:08:00.230 it translates the answers back into standard 00:08:00.230 --> 00:08:04.050 MCP format. It's an abstraction layer. Better 00:08:04.050 --> 00:08:06.370 internal access. Anything else? Yes. A really 00:08:06.370 --> 00:08:08.930 important one. Enhanced security and governance. 00:08:09.470 --> 00:08:11.990 Because communication flows through this standard 00:08:11.990 --> 00:08:14.769 protocol and defined servers, you can implement 00:08:14.769 --> 00:08:17.449 security controls centrally. An administrator 00:08:17.449 --> 00:08:21.310 can configure the MCP server for, say, the company's 00:08:21.310 --> 00:08:24.269 financial database, to allow an AI read -only 00:08:24.269 --> 00:08:26.910 access for analysis but strictly deny it permission 00:08:26.910 --> 00:08:29.769 to execute trades or modify records. Getting 00:08:29.769 --> 00:08:32.549 that kind of fine -grained control is much, much 00:08:32.549 --> 00:08:36.269 harder when you have dozens of ad hoc, one -off 00:08:36.269 --> 00:08:38.649 API integrations everywhere. Right. Centralized 00:08:38.649 --> 00:08:40.889 control points. That makes perfect sense. So 00:08:40.889 --> 00:08:43.110 putting it all together, MCP isn't just about 00:08:43.110 --> 00:08:45.950 making connections easier. It feels like it fundamentally 00:08:45.950 --> 00:08:49.480 changes what AI is. It moves it from being a 00:08:49.480 --> 00:08:51.620 chat partner. To being an autonomous worker, 00:08:51.799 --> 00:08:53.720 an engine that can actually orchestrate workflows 00:08:53.720 --> 00:08:56.120 across different digital tools. It takes requests 00:08:56.120 --> 00:08:58.399 that used to require multiple manual steps and 00:08:58.399 --> 00:09:00.820 just handles them. You use the analogy of asking 00:09:00.820 --> 00:09:03.120 for a recipe versus having a chef cook the meal. 00:09:03.340 --> 00:09:05.340 Exactly. And not just cook the meal, but maybe 00:09:05.340 --> 00:09:06.899 clean up the kitchen and order the groceries 00:09:06.899 --> 00:09:08.639 for next time, too. It's about handling the whole 00:09:08.639 --> 00:09:11.179 process. OK, let's make this concrete. Give us 00:09:11.179 --> 00:09:13.720 that overdue project task example again. But 00:09:13.720 --> 00:09:16.000 walk us through the difference. First, the old 00:09:16.000 --> 00:09:19.559 way, without MCP. Right. The old way, you'd prompt 00:09:19.559 --> 00:09:22.559 your AI. Give me a template for a follow -up 00:09:22.559 --> 00:09:25.039 email to a team member about an overdue project 00:09:25.039 --> 00:09:28.700 task. And you get back? Text? Just the template. 00:09:28.980 --> 00:09:30.799 You still have to find the overdue tasks, find 00:09:30.799 --> 00:09:33.360 the email address, copy, paste, tweak the email, 00:09:33.639 --> 00:09:37.059 save it as a draft. All manual. Precisely. Now 00:09:37.059 --> 00:09:40.659 the new way. With MCP -enabled tools, your prompt 00:09:40.659 --> 00:09:44.080 could be much more ambitious. Review my Q4 product 00:09:44.080 --> 00:09:47.259 launch project in Asana. Find all tasks overdue 00:09:47.259 --> 00:09:49.600 by more than two business days. For each one, 00:09:49.600 --> 00:09:51.440 get the assigned person's email from our Google 00:09:51.440 --> 00:09:54.120 workspace directory. Then, draft a polite but 00:09:54.120 --> 00:09:56.159 firm follow -up email mentioning the task name 00:09:56.159 --> 00:09:58.620 and due date. Save all these drafts in my Gmail 00:09:58.620 --> 00:10:00.480 drafts folder so I can review them. OK, that's 00:10:00.480 --> 00:10:02.539 a much bigger ask. How does the AI actually pull 00:10:02.539 --> 00:10:04.679 that off behind the scenes? It's an orchestrated 00:10:04.679 --> 00:10:07.899 sequence. The AI, the MCP client. First sends 00:10:07.899 --> 00:10:10.899 a request to the Asana MCP server, lists tasks 00:10:10.899 --> 00:10:13.740 in Project X, filter by overdue status. The Asana 00:10:13.740 --> 00:10:15.759 server does that, sends the list back, the AI 00:10:15.759 --> 00:10:18.440 processes it. Then for each overdue task, it 00:10:18.440 --> 00:10:20.879 sends a request to the Google Workspace MCP server, 00:10:21.279 --> 00:10:23.990 find email for user Y. Google Workspace server 00:10:23.990 --> 00:10:26.809 responds. Then the AI connects to the Gmail MCP 00:10:26.809 --> 00:10:29.429 server, create a new draft email with this subject, 00:10:29.590 --> 00:10:31.970 this body, to this recipient. It does this for 00:10:31.970 --> 00:10:34.309 each overdue task. Finally, it might send you 00:10:34.309 --> 00:10:36.289 a message back saying, OK, I've drafted those 00:10:36.289 --> 00:10:38.210 three follow -up emails. They're in your drafts 00:10:38.210 --> 00:10:40.570 folder. All the connections, day lookups, and 00:10:40.570 --> 00:10:43.629 actions are handled via MCP. That is genuinely 00:10:43.629 --> 00:10:45.929 powerful. It's taking maybe, what, 15, 20 minutes 00:10:45.929 --> 00:10:48.470 of annoying admin work and just doing it. What 00:10:48.470 --> 00:10:51.429 other kinds of complex, multi -step things become 00:10:51.429 --> 00:10:54.299 possible? Oh, the possibilities really open up. 00:10:54.480 --> 00:10:56.279 Think about financial analysis and reporting. 00:10:56.460 --> 00:10:59.200 You could say, access QuickBooks via its MCP 00:10:59.200 --> 00:11:02.120 server, pull last quarter's P &L. Now, use the 00:11:02.120 --> 00:11:05.000 AlphaVantage MCP server to get the stock performance 00:11:05.000 --> 00:11:07.399 for our top three competitors over the same period. 00:11:07.779 --> 00:11:10.159 Combine this, make some charts comparing our 00:11:10.159 --> 00:11:12.899 P &L to their stock trends, and save it as a 00:11:12.899 --> 00:11:15.779 report called Q2CompetitiveAnalysis .docs in 00:11:15.779 --> 00:11:18.419 our Google Drive executive reports folder. Wow. 00:11:18.779 --> 00:11:21.500 QuickBooks? A stock data service? Google Drive, 00:11:21.820 --> 00:11:24.000 all working together from one request. Exactly. 00:11:24.340 --> 00:11:27.019 Or consider automated customer support triage. 00:11:27.360 --> 00:11:29.899 Imagine this running constantly. Monitor our 00:11:29.899 --> 00:11:32.299 intercom inbox. If a new ticket comes in with 00:11:32.299 --> 00:11:34.580 keywords like refund request or billing error, 00:11:34.860 --> 00:11:37.139 automatically do this. One, create a high priority 00:11:37.139 --> 00:11:39.340 ticket in JIRA under the Effion billing project. 00:11:39.679 --> 00:11:42.840 Two, assign it to the finance team. Three, post 00:11:42.840 --> 00:11:44.960 a Slack message and hashtag urgent support alerts 00:11:44.960 --> 00:11:47.039 with the customer name and a link to the new 00:11:47.039 --> 00:11:49.600 Jira ticket. So it's proactively monitoring and 00:11:49.600 --> 00:11:51.419 taking multi -step actions based on triggers. 00:11:51.720 --> 00:11:53.740 That's way beyond just responding to a prompt. 00:11:54.019 --> 00:11:56.700 It's a fundamental shift. You state the high 00:11:56.700 --> 00:11:59.639 -level goal, the desired outcome, and the AI, 00:12:00.220 --> 00:12:03.320 armed with MCP connectors, figures out and executes 00:12:03.320 --> 00:12:05.759 the sequence of digital tasks across different 00:12:05.759 --> 00:12:08.559 platforms to get it done. You're managing outcomes, 00:12:08.879 --> 00:12:11.240 not individual clicks. OK, I think people listening 00:12:11.240 --> 00:12:13.259 are probably getting excited about this. So how 00:12:13.259 --> 00:12:15.460 can you, the listener, actually start using this? 00:12:15.519 --> 00:12:18.220 Is it just for developers, or can regular users 00:12:18.220 --> 00:12:20.590 get involved? That's the great thing about a 00:12:20.590 --> 00:12:22.769 growing ecosystem. There are entry points for 00:12:22.769 --> 00:12:25.610 everyone. For most users, the let's call them 00:12:25.610 --> 00:12:27.610 no coders, you're not building anything. You're 00:12:27.610 --> 00:12:30.629 just connecting existing pieces. The absolute 00:12:30.629 --> 00:12:33.610 easiest way in is through native built -in connectors. 00:12:34.350 --> 00:12:37.029 AI platforms like ChatGPT and especially Claude 00:12:37.159 --> 00:12:39.720 are rapidly adding MCP support for common apps 00:12:39.720 --> 00:12:42.159 right into their interface. You might see options 00:12:42.159 --> 00:12:45.179 to connect Google Drive, GitHub, Slack. Claude's 00:12:45.179 --> 00:12:47.659 been particularly active here with really capable 00:12:47.659 --> 00:12:50.159 connectors for Gmail reading, drafting, sending, 00:12:50.639 --> 00:12:52.700 and even things like Claude for Mac or Windows, 00:12:52.960 --> 00:12:55.279 which lets the AI interact with your local computer 00:12:55.279 --> 00:12:57.460 files and apps. Those native ones are usually 00:12:57.460 --> 00:12:59.919 the simplest and safest place to start. So, look 00:12:59.919 --> 00:13:02.360 inside your AI tool first for built -in connections. 00:13:02.669 --> 00:13:05.370 What if the tool you need isn't built in yet? 00:13:05.590 --> 00:13:08.429 Then you look for official third -party MCP servers. 00:13:08.850 --> 00:13:11.250 Major SaaS companies think Notion, Microsoft 00:13:11.250 --> 00:13:14.629 365, Salesforce are increasingly building and 00:13:14.629 --> 00:13:17.370 offering their own official MCP servers. You 00:13:17.370 --> 00:13:19.509 can usually find lists of these on places like 00:13:19.509 --> 00:13:22.789 the official MCP GitHub repository. Setting them 00:13:22.789 --> 00:13:24.669 up is often pretty straightforward for a power 00:13:24.669 --> 00:13:27.309 user. Typically, you go to the service like Notion, 00:13:27.669 --> 00:13:30.250 generate a special API key or token. Then you 00:13:30.250 --> 00:13:33.480 go into your AI tool settings, spot for MCP connections, 00:13:33.720 --> 00:13:35.559 paste in that key and maybe a server address, 00:13:35.940 --> 00:13:37.899 and restart. It's usually manageable without 00:13:37.899 --> 00:13:39.980 coding. Okay, official servers from the companies 00:13:39.980 --> 00:13:42.409 themselves. Are there other options? Yes, there 00:13:42.409 --> 00:13:45.250 are community connectors. Because MCP is open 00:13:45.250 --> 00:13:47.009 source, the community is building connectors 00:13:47.009 --> 00:13:49.250 for all sorts of niche tools. You might find 00:13:49.250 --> 00:13:51.769 these on websites dedicated to MCP, like mcp 00:13:51.769 --> 00:13:54.929 .sosa is one I've seen. But a really important 00:13:54.929 --> 00:13:57.750 word of caution here. While many community connectors 00:13:57.750 --> 00:14:00.389 are fantastic, built by talented developers, 00:14:00.549 --> 00:14:02.929 some might be buggy or might stop being updated. 00:14:03.409 --> 00:14:05.850 Or, in the worst case, could even be malicious, 00:14:06.009 --> 00:14:08.470 designed to steal your data or credentials. Right, 00:14:08.470 --> 00:14:10.830 like browser extensions or apps from unknown 00:14:10.830 --> 00:14:14.720 sources. Use with care. Exactly. Use community 00:14:14.720 --> 00:14:17.240 connectors for less critical tasks first. Test 00:14:17.240 --> 00:14:19.799 them out. And definitely think twice before giving 00:14:19.799 --> 00:14:22.259 one access to highly sensitive data like your 00:14:22.259 --> 00:14:25.059 email or financial accounts. Stick to official 00:14:25.059 --> 00:14:26.919 or native connectors for the really important 00:14:26.919 --> 00:14:29.549 stuff if you can. Good advice. Okay, so that 00:14:29.549 --> 00:14:31.350 covers users. What about developers who want 00:14:31.350 --> 00:14:34.590 to build with this? For developers, MCP is incredibly 00:14:34.590 --> 00:14:37.230 powerful. You can build those private MCP servers 00:14:37.230 --> 00:14:39.370 we talked about, giving your AI applications 00:14:39.370 --> 00:14:41.929 secure access to your company's unique internal 00:14:41.929 --> 00:14:45.490 tools and data. Imagine an AI interface for your 00:14:45.490 --> 00:14:49.029 custom CRM or inventory system. The MCP standard 00:14:49.029 --> 00:14:51.230 itself is well -documented, and there are official 00:14:51.230 --> 00:14:53.250 libraries available for popular languages like 00:14:53.250 --> 00:14:55.850 Python, TypeScript, Java, making it easier to 00:14:55.850 --> 00:14:58.909 get started. great learning resources popping 00:14:58.909 --> 00:15:01.470 up, like courses from deeplearning .ai, specifically 00:15:01.470 --> 00:15:04.210 on building these kinds of rich context AI apps 00:15:04.210 --> 00:15:06.750 with Anthropix tools. When you actually go to 00:15:06.750 --> 00:15:08.490 build an MCP server, there are generally two 00:15:08.490 --> 00:15:10.610 ways to approach it. You can do a low -level 00:15:10.610 --> 00:15:12.809 implementation, basically coding directly against 00:15:12.809 --> 00:15:15.889 the MCP protocol specifications. This gives you 00:15:15.889 --> 00:15:18.269 maximum control but requires a deeper understanding 00:15:18.269 --> 00:15:20.710 of the protocol details. Or you can use high 00:15:20.710 --> 00:15:22.950 -level frameworks, libraries like Fast MCP for 00:15:22.950 --> 00:15:25.509 Python or Easy MCP for TypeScript, Abstract, 00:15:25.529 --> 00:15:27.860 away a lot of the protocol complexity. You can 00:15:27.860 --> 00:15:29.860 often just define your tools as regular functions 00:15:29.860 --> 00:15:32.220 in your code, add some decorators, and the framework 00:15:32.220 --> 00:15:34.639 handles the MCP communication layer for you. 00:15:35.019 --> 00:15:37.440 Much faster development. So options for both 00:15:37.440 --> 00:15:40.440 deep control and rapid development. It sounds 00:15:40.440 --> 00:15:42.580 like the ecosystem is really starting to mature. 00:15:43.039 --> 00:15:45.240 We should acknowledge, though, MCP is still pretty 00:15:45.240 --> 00:15:47.700 new, right? It's not perfect yet. Oh, absolutely. 00:15:47.820 --> 00:15:50.230 It's definitely a young standard. Like any emerging 00:15:50.230 --> 00:15:52.850 tech, it has its rough edges. Some connectors 00:15:52.850 --> 00:15:54.889 might still be a bit buggy or limited in what 00:15:54.889 --> 00:15:58.029 they can do. Setting up a local development environment 00:15:58.029 --> 00:16:01.389 can sometimes be fiddly. And maybe most importantly, 00:16:01.769 --> 00:16:05.029 the security implications of giving AI autonomous 00:16:05.029 --> 00:16:07.990 agency to act on your behalf across multiple 00:16:07.990 --> 00:16:11.429 systems. Well, those are profound. We need... 00:16:11.289 --> 00:16:14.210 really robust security models, granular permissions, 00:16:14.590 --> 00:16:17.549 clear audit trails. It's all still evolving and 00:16:17.549 --> 00:16:19.710 requires very careful thought. Definitely something 00:16:19.710 --> 00:16:21.730 to keep a close eye on. But even with those growing 00:16:21.730 --> 00:16:24.129 pains, the direction seems pretty clear. MCP, 00:16:24.129 --> 00:16:26.230 or something very much like it, looks set to 00:16:26.230 --> 00:16:28.950 become as fundamental to the AI stack as, say, 00:16:29.450 --> 00:16:31.529 REST APIs became for the web. It's that missing 00:16:31.529 --> 00:16:33.850 link, that standardized way for intelligence 00:16:33.850 --> 00:16:36.169 to actually translate into action in the digital 00:16:36.169 --> 00:16:39.179 world. For you listening, whether you're using 00:16:39.179 --> 00:16:41.879 AI every day at work or you're a developer building 00:16:41.879 --> 00:16:44.879 AI apps or a business leader thinking about strategy, 00:16:45.299 --> 00:16:47.879 now really is the time to start exploring. See 00:16:47.879 --> 00:16:49.960 what MCP connectors are available for the tools 00:16:49.960 --> 00:16:52.679 you rely on. Getting familiar with this, maybe 00:16:52.679 --> 00:16:54.519 even contributing back if you're a developer, 00:16:55.159 --> 00:16:57.919 feels less like an option and more like a strategic 00:16:57.919 --> 00:17:00.419 necessity for the future. Couldn't agree more. 00:17:00.600 --> 00:17:03.240 The era of the AI just being a standalone chatbot 00:17:03.240 --> 00:17:06.460 in a window. That's ending the age of the integrated 00:17:06.460 --> 00:17:09.259 automated AI powerhouse that works across your 00:17:09.259 --> 00:17:11.740 digital landscape. That's really just beginning. 00:17:12.180 --> 00:17:13.799 So here's something to think about. Looking at 00:17:13.799 --> 00:17:16.119 your own work, your own daily digital tasks. 00:17:16.759 --> 00:17:19.680 What complex multi -step process, something tedious 00:17:19.680 --> 00:17:22.180 or time consuming today, could you imagine an 00:17:22.180 --> 00:17:24.720 AI orchestrating for you using this kind of technology? 00:17:25.359 --> 00:17:27.559 What could you achieve if that time was freed 00:17:27.559 --> 00:17:27.779 up?