1 00:00:00,000 --> 00:00:02,220 Today's topic is one that almost everyone has heard of, 2 00:00:02,220 --> 00:00:03,720 but few can explain clearly. 3 00:00:03,720 --> 00:00:07,280 MCP, you've probably seen it called HTTP for AI agents. 4 00:00:07,280 --> 00:00:08,760 But what does that actually mean? 5 00:00:08,760 --> 00:00:10,400 By the end of this episode, you'll understand 6 00:00:10,400 --> 00:00:12,760 what MCP is, what Azure MCP server does, 7 00:00:12,760 --> 00:00:15,520 and why it matters for how AI tools connect to your data. 8 00:00:15,520 --> 00:00:17,360 We'll break it down into the major building blocks, 9 00:00:17,360 --> 00:00:18,720 starting with the big picture, 10 00:00:18,720 --> 00:00:21,400 then the protocol itself, then Microsoft's implementation, 11 00:00:21,400 --> 00:00:24,280 and finally a real example you can actually use. 12 00:00:24,280 --> 00:00:26,320 The big picture, why this matters now? 13 00:00:26,320 --> 00:00:28,120 So here's the situation we're in right now. 14 00:00:28,120 --> 00:00:30,920 Everyone is building AI agents, not just tech companies, 15 00:00:30,920 --> 00:00:34,000 but banks, hospitals, manufacturers, even small businesses. 16 00:00:34,000 --> 00:00:35,720 They're all trying to build AI assistants 17 00:00:35,720 --> 00:00:38,000 that can actually do things, not just answer questions, 18 00:00:38,000 --> 00:00:39,240 but take real action. 19 00:00:39,240 --> 00:00:41,720 Update a record, check inventory, deploy code, 20 00:00:41,720 --> 00:00:43,320 or fix a broken server. 21 00:00:43,320 --> 00:00:44,800 And here's where the problem shows up. 22 00:00:44,800 --> 00:00:48,080 These agents need to connect to real tools and real data. 23 00:00:48,080 --> 00:00:52,160 Databases, storage accounts, ticketing systems, CRM platforms. 24 00:00:52,160 --> 00:00:54,360 The way most people do this today is the hard way. 25 00:00:54,360 --> 00:00:56,000 They write custom integration code 26 00:00:56,000 --> 00:00:58,640 for every single tool, every API, every data source. 27 00:00:58,640 --> 00:01:01,040 One connector for the database, another for file storage, 28 00:01:01,040 --> 00:01:02,320 a third for the monitoring system. 29 00:01:02,320 --> 00:01:05,520 Each one needs its own authentication, its own error handling, 30 00:01:05,520 --> 00:01:08,320 its own way of formatting requests and responses. 31 00:01:08,320 --> 00:01:11,160 This approaches fragile, time-consuming, and it doesn't scale. 32 00:01:11,160 --> 00:01:14,040 Every AI assistant you build needs its own pile of glue code 33 00:01:14,040 --> 00:01:15,760 to talk to the services it needs. 34 00:01:15,760 --> 00:01:17,640 Switch from one AI platform to another, 35 00:01:17,640 --> 00:01:19,560 and you rebuild everything from scratch. 36 00:01:19,560 --> 00:01:22,280 Most AI demos you see work great in isolation. 37 00:01:22,280 --> 00:01:25,400 They answer questions, generate text, look impressive. 38 00:01:25,400 --> 00:01:27,160 At the moment they need to touch a real system, 39 00:01:27,160 --> 00:01:28,000 they fall apart. 40 00:01:28,000 --> 00:01:29,440 That's where the complexity lives. 41 00:01:29,440 --> 00:01:30,680 MCP changes this. 42 00:01:30,680 --> 00:01:32,400 It gives AI a standard way to reach out 43 00:01:32,400 --> 00:01:33,680 and grab what it needs. 44 00:01:33,680 --> 00:01:36,400 One protocol, one way of talking to tools. 45 00:01:36,400 --> 00:01:39,960 Any AI assistant that speaks MCP can use any MCP server. 46 00:01:39,960 --> 00:01:41,560 It's like the difference between having 47 00:01:41,560 --> 00:01:44,440 a separate remote control for every device in your house 48 00:01:44,440 --> 00:01:47,200 versus one universal remote that works with everything. 49 00:01:47,200 --> 00:01:49,400 So let's start with the simplest definition. 50 00:01:49,400 --> 00:01:51,520 What is MCP, the protocol explained? 51 00:01:51,520 --> 00:01:52,600 What exactly is MCP? 52 00:01:52,600 --> 00:01:53,880 Let me explain it in plain English. 53 00:01:53,880 --> 00:01:56,320 MCP stands for model context protocol, 54 00:01:56,320 --> 00:01:58,640 and it's an open standard that tells AI agents 55 00:01:58,640 --> 00:02:01,840 how to talk to external tools, APIs, and data sources. 56 00:02:01,840 --> 00:02:03,560 Think of it as a common language for AI. 57 00:02:03,560 --> 00:02:06,720 Before MCP, every AI assistant had to learn a different language 58 00:02:06,720 --> 00:02:09,120 for every tool it wanted to use. 59 00:02:09,120 --> 00:02:11,240 MCP says, let's all agree on one language, 60 00:02:11,240 --> 00:02:14,000 and then every tool just needs to speak that one language. 61 00:02:14,000 --> 00:02:16,920 Under the hood, it's based on Jason RPC 2.0, 62 00:02:16,920 --> 00:02:18,840 a simple message format for sending requests 63 00:02:18,840 --> 00:02:20,000 and getting responses. 64 00:02:20,000 --> 00:02:22,360 Nothing fancy, and it's been around for years. 65 00:02:22,360 --> 00:02:24,240 What MCP adds on top is structure. 66 00:02:24,240 --> 00:02:26,720 It defines three key things an AI agent can do 67 00:02:26,720 --> 00:02:30,120 with any MCP server, tools, resources, and prompts. 68 00:02:30,120 --> 00:02:31,720 Tools are actions the AI can take 69 00:02:31,720 --> 00:02:34,000 like creating a record or running a query. 70 00:02:34,000 --> 00:02:36,360 Resources are pieces of data the AI can read, 71 00:02:36,360 --> 00:02:38,480 like a customer profile or a log file, 72 00:02:38,480 --> 00:02:41,440 and prompts are templates that guide how the AI should behave 73 00:02:41,440 --> 00:02:42,840 in certain situations. 74 00:02:42,840 --> 00:02:44,200 Here's the simplest way to think about it. 75 00:02:44,200 --> 00:02:46,480 Imagine you have a universal remote control 76 00:02:46,480 --> 00:02:48,400 that remote can talk to your TV, your sound system, 77 00:02:48,400 --> 00:02:49,880 your streaming box, your lights. 78 00:02:49,880 --> 00:02:53,320 One remote, many devices, MCP is that remote for AI agents. 79 00:02:53,320 --> 00:02:56,080 One protocol, many tools, you build one MCP server 80 00:02:56,080 --> 00:02:57,800 that exposes your tools and data, 81 00:02:57,800 --> 00:03:00,760 and any MCP compatible AI agent can use it. 82 00:03:00,760 --> 00:03:03,680 GitHub co-pilot, Claude, chatGPT, custom agents, 83 00:03:03,680 --> 00:03:05,280 they all speak the same protocol, 84 00:03:05,280 --> 00:03:07,520 and this isn't a proprietary Microsoft thing. 85 00:03:07,520 --> 00:03:09,520 MCP was originally created by Anthropic, 86 00:03:09,520 --> 00:03:10,920 the company behind Claude, 87 00:03:10,920 --> 00:03:13,280 but it's been adopted by OpenAI, Google, Microsoft, 88 00:03:13,280 --> 00:03:14,120 and many others. 89 00:03:14,120 --> 00:03:16,200 It's becoming the standard way for AI agents 90 00:03:16,200 --> 00:03:17,880 to connect to the outside world. 91 00:03:17,880 --> 00:03:21,320 Not a vendor lock-in, but an open standard everyone can use. 92 00:03:21,320 --> 00:03:23,120 The problem MCP solves. 93 00:03:23,120 --> 00:03:24,560 Here's the old way of doing things. 94 00:03:24,560 --> 00:03:27,720 Before MCP, every AI integration was a custom project. 95 00:03:27,720 --> 00:03:30,360 Say you want your AI assistant to check inventory levels 96 00:03:30,360 --> 00:03:31,480 in your database. 97 00:03:31,480 --> 00:03:33,240 Great, now you need to write code. 98 00:03:33,240 --> 00:03:34,920 You need to figure out authentication. 99 00:03:34,920 --> 00:03:37,400 How does the AI prove it's allowed to read that data? 100 00:03:37,400 --> 00:03:39,000 You need to handle errors. 101 00:03:39,000 --> 00:03:40,880 What happens when the database is down? 102 00:03:40,880 --> 00:03:42,120 You need to format the response. 103 00:03:42,120 --> 00:03:44,400 How does the database output get turned into something 104 00:03:44,400 --> 00:03:46,000 the AI can understand? 105 00:03:46,000 --> 00:03:48,680 Weeks of development for one connection to one database. 106 00:03:48,680 --> 00:03:50,120 That's the integration tax. 107 00:03:50,120 --> 00:03:52,920 One AI assistant, one database, weeks of work. 108 00:03:52,920 --> 00:03:55,720 Now multiply that by every tool your AI needs to touch. 109 00:03:55,720 --> 00:03:58,080 The CRM, the ticketing system, the file storage, 110 00:03:58,080 --> 00:04:01,040 the monitoring logs, each one requires its own custom code, 111 00:04:01,040 --> 00:04:03,640 its own authentication pattern, its own error handling, 112 00:04:03,640 --> 00:04:05,640 and none of that code works with any other tool. 113 00:04:05,640 --> 00:04:07,600 You're building the same thing over and over 114 00:04:07,600 --> 00:04:09,280 just slightly different each time. 115 00:04:09,280 --> 00:04:10,680 And here's the security problem. 116 00:04:10,680 --> 00:04:13,040 Every custom integration has its own authentication. 117 00:04:13,040 --> 00:04:15,320 Some use API keys, some use OAuth, 118 00:04:15,320 --> 00:04:18,040 some use certificates, some use basic OAuth. 119 00:04:18,040 --> 00:04:19,880 You end up with a dozen different ways for your AI 120 00:04:19,880 --> 00:04:23,080 to access data and no consistent way to govern any of it. 121 00:04:23,080 --> 00:04:23,800 Who has access? 122 00:04:23,800 --> 00:04:24,760 What can they do? 123 00:04:24,760 --> 00:04:25,800 Where are the logs? 124 00:04:25,800 --> 00:04:27,960 Most organizations can't answer those questions 125 00:04:27,960 --> 00:04:30,360 for their AI integrations because the integrations 126 00:04:30,360 --> 00:04:33,200 were built by different teams at different times 127 00:04:33,200 --> 00:04:34,520 with different standards. 128 00:04:34,520 --> 00:04:35,960 The fragmentation problem is real. 129 00:04:35,960 --> 00:04:38,720 Let's say you build your AI assistant on open AI. 130 00:04:38,720 --> 00:04:40,960 You write all your integrations for open AI's function 131 00:04:40,960 --> 00:04:41,960 calling format. 132 00:04:41,960 --> 00:04:43,760 Then six months later, your company 133 00:04:43,760 --> 00:04:46,480 decides to switch to Claude or you want to use both. 134 00:04:46,480 --> 00:04:48,160 Now you rebuild every single integration. 135 00:04:48,160 --> 00:04:49,920 Same tools, same data, different formats. 136 00:04:49,920 --> 00:04:51,680 That's not progress, that's busy work. 137 00:04:51,680 --> 00:04:53,240 The real cost adds up fast. 138 00:04:53,240 --> 00:04:56,400 Most large organizations already run dozens of AI agents. 139 00:04:56,400 --> 00:04:59,760 According to IDC, the average enterprise runs 47 of them. 140 00:04:59,760 --> 00:05:02,200 Each one has its own pile of custom connectors. 141 00:05:02,200 --> 00:05:03,760 And here's a stat that might surprise you. 142 00:05:03,760 --> 00:05:08,040 68% of CIOs can't even report total AI agents spend. 143 00:05:08,040 --> 00:05:10,040 They don't know what they're spending on these integrations 144 00:05:10,040 --> 00:05:13,280 because it's scattered across teams, projects, and budgets. 145 00:05:13,280 --> 00:05:14,640 MCP flips this whole model. 146 00:05:14,640 --> 00:05:17,120 Instead of writing custom code for every AI tool, 147 00:05:17,120 --> 00:05:19,160 you build one MCP server. 148 00:05:19,160 --> 00:05:20,880 That server exposes your tools and data 149 00:05:20,880 --> 00:05:22,280 through a standard interface. 150 00:05:22,280 --> 00:05:24,640 Then any MCP compatible agent can use it. 151 00:05:24,640 --> 00:05:28,560 Co-pilot, Claude, chatGPT, custom agents, your own internal tools. 152 00:05:28,560 --> 00:05:30,080 Write once, use everywhere. 153 00:05:30,080 --> 00:05:31,120 That's the shift. 154 00:05:31,120 --> 00:05:32,840 What is Azure MCP server? 155 00:05:32,840 --> 00:05:34,800 So we've covered the problem MCP solves. 156 00:05:34,800 --> 00:05:36,920 Now let's talk about Microsoft's answer to it. 157 00:05:36,920 --> 00:05:40,240 The Azure MCP server is Microsoft's official take on MCP. 158 00:05:40,240 --> 00:05:42,440 Think of it as a bridge that lets AI agents 159 00:05:42,440 --> 00:05:45,720 securely reach into your Azure resources and manage them. 160 00:05:45,720 --> 00:05:48,000 And I want to be clear about what this thing actually is. 161 00:05:48,000 --> 00:05:49,400 It's not a separate product you buy. 162 00:05:49,400 --> 00:05:50,760 It's not a new Azure service. 163 00:05:50,760 --> 00:05:52,360 It's a tool, an open source server 164 00:05:52,360 --> 00:05:54,360 that sits right between your AI assistant 165 00:05:54,360 --> 00:05:55,560 and your Azure environment. 166 00:05:55,560 --> 00:05:57,520 It works with GitHub co-pilot agent mode. 167 00:05:57,520 --> 00:05:58,680 It works with semantic kernel. 168 00:05:58,680 --> 00:06:00,040 It works with Claude desktop. 169 00:06:00,040 --> 00:06:03,280 And it works with any custom MCP agent you build yourself. 170 00:06:03,280 --> 00:06:06,000 Any client that speaks MCP can connect to it. 171 00:06:06,000 --> 00:06:08,720 Once connected, that agent gets access to your Azure resources 172 00:06:08,720 --> 00:06:09,920 through plain natural language. 173 00:06:09,920 --> 00:06:12,200 You don't need to know the exact API call. 174 00:06:12,200 --> 00:06:13,920 You don't need to remember the SDK method. 175 00:06:13,920 --> 00:06:14,960 You just ask. 176 00:06:14,960 --> 00:06:16,240 Here's the core idea. 177 00:06:16,240 --> 00:06:17,640 Azure has dozens of services. 178 00:06:17,640 --> 00:06:20,720 Storage, databases, monitoring, configuration, containers. 179 00:06:20,720 --> 00:06:23,360 Each one comes with its own SDK, its own API, 180 00:06:23,360 --> 00:06:25,040 its own authentication pattern. 181 00:06:25,040 --> 00:06:28,040 Azure MCP server wraps all of them into one unified interface, 182 00:06:28,040 --> 00:06:30,360 one server, one protocol, all the tools. 183 00:06:30,360 --> 00:06:31,600 Why did Microsoft build this? 184 00:06:31,600 --> 00:06:34,520 Because the fragmentation problem is especially bad in Azure. 185 00:06:34,520 --> 00:06:36,200 Imagine you're building an AI agent 186 00:06:36,200 --> 00:06:38,160 that needs to check Azure Monitor logs, 187 00:06:38,160 --> 00:06:39,960 update an app configuration setting, 188 00:06:39,960 --> 00:06:41,440 and deploy a new container version. 189 00:06:41,440 --> 00:06:44,240 That's three different SDKs, three different authentication 190 00:06:44,240 --> 00:06:46,920 flows, three different ways of handling errors. 191 00:06:46,920 --> 00:06:50,160 Azure MCP server collapses that into one session, 192 00:06:50,160 --> 00:06:52,960 one authentication, one consistent way of working. 193 00:06:52,960 --> 00:06:54,440 Now, here's the important part. 194 00:06:54,440 --> 00:06:55,840 This isn't just for developers. 195 00:06:55,840 --> 00:06:56,880 Yes, developers will use it. 196 00:06:56,880 --> 00:07:00,240 But the real power is that it enables natural language control 197 00:07:00,240 --> 00:07:01,760 of Azure resources. 198 00:07:01,760 --> 00:07:04,240 Someone who doesn't know PowerShell or the Azure CLI 199 00:07:04,240 --> 00:07:06,200 can ask their AI assistant, show me 200 00:07:06,200 --> 00:07:09,880 the last hour of errors for the payment service, and get an answer. 201 00:07:09,880 --> 00:07:11,840 The MCP server handles the translation. 202 00:07:11,840 --> 00:07:13,280 It turns the natural language request 203 00:07:13,280 --> 00:07:16,680 into the appropriate API call, runs it against Azure Monitor, 204 00:07:16,680 --> 00:07:18,040 and returns the result. 205 00:07:18,040 --> 00:07:22,160 As of today, Azure MCP server is built into Visual Studio 2026. 206 00:07:22,160 --> 00:07:23,600 It's open source on GitHub. 207 00:07:23,600 --> 00:07:26,560 It's actively maintained by Microsoft's Azure SDK team, 208 00:07:26,560 --> 00:07:29,320 and it's in public preview, meaning you can use it right now 209 00:07:29,320 --> 00:07:31,880 with more capabilities being added regularly. 210 00:07:31,880 --> 00:07:33,760 Core services, you can access. 211 00:07:33,760 --> 00:07:36,480 So what can you actually do with Azure MCP server? 212 00:07:36,480 --> 00:07:38,640 Let's look at the specific services it can reach. 213 00:07:38,640 --> 00:07:41,280 The list is growing, but here's what's supported right now. 214 00:07:41,280 --> 00:07:42,280 Azure Storage. 215 00:07:42,280 --> 00:07:44,960 You can list storage accounts, manage blob containers, 216 00:07:44,960 --> 00:07:47,240 query tables, all through natural language. 217 00:07:47,240 --> 00:07:49,360 Instead of writing a script to check your storage, 218 00:07:49,360 --> 00:07:50,520 you just ask. 219 00:07:50,520 --> 00:07:53,080 Show me the containers in my production storage account. 220 00:07:53,080 --> 00:07:56,320 The MCP server handles the API call, formats the response, 221 00:07:56,320 --> 00:07:57,600 and gives you the answer. 222 00:07:57,600 --> 00:07:58,800 Azure Cosmos DB. 223 00:07:58,800 --> 00:07:59,600 This one is huge. 224 00:07:59,600 --> 00:08:01,680 You can query databases, manage containers, 225 00:08:01,680 --> 00:08:04,320 and run SQL queries against your no-sequel data. 226 00:08:04,320 --> 00:08:06,680 If you're building an AI agent that needs to look up customer 227 00:08:06,680 --> 00:08:08,240 records or order histories, 228 00:08:08,240 --> 00:08:10,680 this is how it gets that data, no custom connector, 229 00:08:10,680 --> 00:08:14,560 no SDK code, just a two call through MCP as your monitor. 230 00:08:14,560 --> 00:08:16,400 This is where the real power shows up. 231 00:08:16,400 --> 00:08:19,440 You can query logs using KQL, Kusto query language, 232 00:08:19,440 --> 00:08:21,640 without writing a single KQL command yourself. 233 00:08:21,640 --> 00:08:24,440 List workspaces, check available tables, 234 00:08:24,440 --> 00:08:26,920 ask, show me the last 24 hours of errors, 235 00:08:26,920 --> 00:08:29,880 and the MCP server translates that into the right query, 236 00:08:29,880 --> 00:08:32,720 runs it against log analytics, and returns the results. 237 00:08:32,720 --> 00:08:34,880 For anyone who's ever struggled with KQL syntax, 238 00:08:34,880 --> 00:08:37,120 this alone is worth the price of admission. 239 00:08:37,120 --> 00:08:38,400 Azure app configuration. 240 00:08:38,400 --> 00:08:40,840 You can manage key value pairs, handle labeled configurations, 241 00:08:40,840 --> 00:08:42,120 lock and unlock settings. 242 00:08:42,120 --> 00:08:44,680 If your application reads configuration from Azure app 243 00:08:44,680 --> 00:08:49,320 configuration, your AI agent can check those values, update them, 244 00:08:49,320 --> 00:08:52,080 and verify changes all through conversation. 245 00:08:52,080 --> 00:08:53,080 And it's not just services. 246 00:08:53,080 --> 00:08:55,160 The Azure CLI and Azure developer CLI, 247 00:08:55,160 --> 00:08:57,440 ACD are also exposed as tools. 248 00:08:57,440 --> 00:09:00,840 That means your AI agent can run any Azure CLI command directly. 249 00:09:00,840 --> 00:09:04,680 It can provision resources, deploy applications, manage infrastructure, 250 00:09:04,680 --> 00:09:06,440 the same commands you type into a terminal, 251 00:09:06,440 --> 00:09:09,200 but triggered by natural language through an AI assistant. 252 00:09:09,200 --> 00:09:11,680 What this means in practice is that one MCP server 253 00:09:11,680 --> 00:09:14,880 can replace dozens of SDKs and API calls. 254 00:09:14,880 --> 00:09:17,120 You don't need to install the Azure storage SDK, 255 00:09:17,120 --> 00:09:19,400 the Cosmos DB SDK, the monitor SDK, 256 00:09:19,400 --> 00:09:21,400 and learn three different authentication patterns. 257 00:09:21,400 --> 00:09:23,960 You install one MCP server, authenticate ones, 258 00:09:23,960 --> 00:09:25,600 and all those capabilities are available 259 00:09:25,600 --> 00:09:26,840 through a single interface. 260 00:09:26,840 --> 00:09:29,240 Think about a typical troubleshooting scenario. 261 00:09:29,240 --> 00:09:31,000 Your application is throwing errors. 262 00:09:31,000 --> 00:09:33,200 You need to check the logs, look at the configuration, 263 00:09:33,200 --> 00:09:35,120 verify the storage account is accessible, 264 00:09:35,120 --> 00:09:36,720 and check the database connection. 265 00:09:36,720 --> 00:09:38,640 Before MCP, that's four different tools, 266 00:09:38,640 --> 00:09:41,520 four different logins, four different ways of getting information. 267 00:09:41,520 --> 00:09:44,400 With Azure MCP server, it's one conversation. 268 00:09:44,400 --> 00:09:47,160 Show me the errors, check my config, verify my storage, 269 00:09:47,160 --> 00:09:48,840 and test my database connection. 270 00:09:48,840 --> 00:09:50,560 The server handles all of it. 271 00:09:50,560 --> 00:09:51,880 How it actually works? 272 00:09:51,880 --> 00:09:53,080 Let's look under the hood. 273 00:09:53,080 --> 00:09:55,120 How does this actually work behind the scenes? 274 00:09:55,120 --> 00:09:57,040 Authentication uses EntraID, 275 00:09:57,040 --> 00:09:58,760 that's your Azure Active Directory, 276 00:09:58,760 --> 00:10:00,560 through the Azure Identity Library. 277 00:10:00,560 --> 00:10:02,360 When your AI agent makes a request, 278 00:10:02,360 --> 00:10:04,840 it authenticates using your Azure Identity. 279 00:10:04,840 --> 00:10:07,360 The same credentials you use to log into the Azure Portal, 280 00:10:07,360 --> 00:10:09,520 no separate API keys, no shared service accounts, 281 00:10:09,520 --> 00:10:10,720 just your identity. 282 00:10:10,720 --> 00:10:12,320 Think of it like a reception desk. 283 00:10:12,320 --> 00:10:14,600 The server checks your ID before letting you in. 284 00:10:14,600 --> 00:10:17,000 Now, authorization respects Azure R-Back, 285 00:10:17,000 --> 00:10:18,800 role-based access control. 286 00:10:18,800 --> 00:10:21,560 The MCP server doesn't grant any new permissions. 287 00:10:21,560 --> 00:10:24,400 It only allows what your account already has permission to do. 288 00:10:24,400 --> 00:10:26,880 If you can read storage accounts in the Portal, 289 00:10:26,880 --> 00:10:28,520 the MCP server can read them. 290 00:10:28,520 --> 00:10:29,960 If you can't delete databases, 291 00:10:29,960 --> 00:10:31,400 the MCP server can't either. 292 00:10:31,400 --> 00:10:33,560 It inherits Azure's existing security model. 293 00:10:33,560 --> 00:10:34,920 It doesn't add new attack service. 294 00:10:34,920 --> 00:10:37,600 So if you already trust Azure to run your applications, 295 00:10:37,600 --> 00:10:39,560 you already trust the security model here. 296 00:10:39,560 --> 00:10:41,320 For transport, there are two options. 297 00:10:41,320 --> 00:10:43,920 Streamable HTTP for remote connections. 298 00:10:43,920 --> 00:10:45,880 When your AI assistant is running in the cloud 299 00:10:45,880 --> 00:10:47,280 or on a different machine, 300 00:10:47,280 --> 00:10:49,040 an STDO for local development, 301 00:10:49,040 --> 00:10:50,960 when you're running everything on your own machine, 302 00:10:50,960 --> 00:10:53,880 the Streamable HTTP option is the one Microsoft recommends 303 00:10:53,880 --> 00:10:55,920 for production because it works well 304 00:10:55,920 --> 00:10:57,720 with load balancers and proxies. 305 00:10:57,720 --> 00:10:59,480 Think of it like a phone call for remote 306 00:10:59,480 --> 00:11:01,360 and a direct conversation for local. 307 00:11:01,360 --> 00:11:02,680 Sessions are important here. 308 00:11:02,680 --> 00:11:05,560 The MCP server maintains state during a conversation. 309 00:11:05,560 --> 00:11:08,160 So when your agent asks a question, gets an answer, 310 00:11:08,160 --> 00:11:09,680 and then asks a follow-up question, 311 00:11:09,680 --> 00:11:11,640 the server remembers what happened before. 312 00:11:11,640 --> 00:11:13,480 Tools can build on previous steps. 313 00:11:13,480 --> 00:11:15,960 You can query logs, identify a problem, 314 00:11:15,960 --> 00:11:19,160 create a fix, and deploy it all within one session 315 00:11:19,160 --> 00:11:20,520 without losing context. 316 00:11:20,520 --> 00:11:22,240 It's like having a coffee chat where you pick up 317 00:11:22,240 --> 00:11:23,440 right where you left off. 318 00:11:23,440 --> 00:11:25,560 The server uses Azure SDKs under the hood, 319 00:11:25,560 --> 00:11:27,960 which means it inherits all the built in retry logic 320 00:11:27,960 --> 00:11:30,040 and error handling those SDKs provide. 321 00:11:30,040 --> 00:11:32,480 If a request fails because of a transient network issue, 322 00:11:32,480 --> 00:11:34,120 it retries automatically. 323 00:11:34,120 --> 00:11:37,320 If a service is unavailable, it handles the error gracefully. 324 00:11:37,320 --> 00:11:39,120 You don't have to write any of that yourself. 325 00:11:39,120 --> 00:11:40,040 It's already baked in. 326 00:11:40,040 --> 00:11:41,720 Now the security model uses something called 327 00:11:41,720 --> 00:11:44,160 on behalf of authentication or OBO. 328 00:11:44,160 --> 00:11:45,200 This is important. 329 00:11:45,200 --> 00:11:47,120 When your AI agent calls a tool, 330 00:11:47,120 --> 00:11:49,160 the MCP server passes your identity 331 00:11:49,160 --> 00:11:50,880 through to the downstream service. 332 00:11:50,880 --> 00:11:52,760 The storage account sees your identity, 333 00:11:52,760 --> 00:11:54,160 not the server's identity. 334 00:11:54,160 --> 00:11:55,960 The database sees your identity, 335 00:11:55,960 --> 00:11:57,880 not some shared service account. 336 00:11:57,880 --> 00:12:00,160 This means every action is attributable to you. 337 00:12:00,160 --> 00:12:01,840 Every audit log shows who did what. 338 00:12:01,840 --> 00:12:04,400 No anonymous service accounts doing things nobody can trace. 339 00:12:04,400 --> 00:12:06,280 The server says, I'm here on behalf of Mirko 340 00:12:06,280 --> 00:12:08,840 and the service sees you, not the server. 341 00:12:08,840 --> 00:12:10,960 For enterprises that need even more control, 342 00:12:10,960 --> 00:12:14,040 you can front MCP servers with Azure API management. 343 00:12:14,040 --> 00:12:15,600 This gives you governance policies, 344 00:12:15,600 --> 00:12:18,240 rate limiting audit trails and content safety checks 345 00:12:18,240 --> 00:12:19,480 at the protocol layer. 346 00:12:19,480 --> 00:12:21,400 You can control which tools are available, 347 00:12:21,400 --> 00:12:22,960 who can use them and how often. 348 00:12:22,960 --> 00:12:24,960 It's the same API management you already use 349 00:12:24,960 --> 00:12:27,920 for your REST APIs now applied to MCP tools. 350 00:12:27,920 --> 00:12:29,280 Here's the key takeaway. 351 00:12:29,280 --> 00:12:32,400 As your MCP server doesn't invent new security mechanisms, 352 00:12:32,400 --> 00:12:34,080 it uses the ones Azure already has. 353 00:12:34,080 --> 00:12:35,800 Entra ID for authentication, 354 00:12:35,800 --> 00:12:38,520 R-Back for authorization, SDKs for reliability, 355 00:12:38,520 --> 00:12:40,320 API management for governance. 356 00:12:40,320 --> 00:12:42,680 If you already trust Azure to run your applications, 357 00:12:42,680 --> 00:12:44,160 you already trust the security model 358 00:12:44,160 --> 00:12:46,000 that MCP server uses. 359 00:12:46,000 --> 00:12:47,560 Real-world example in action. 360 00:12:47,560 --> 00:12:49,920 Let's make this real with an example you'd actually encounter. 361 00:12:49,920 --> 00:12:52,360 Imagine you're running a container app on Azure. 362 00:12:52,360 --> 00:12:54,880 You go to the URL and you get a 404 error. 363 00:12:54,880 --> 00:12:55,880 The app won't start. 364 00:12:55,880 --> 00:12:57,640 The Azure portal shows an activation error, 365 00:12:57,640 --> 00:12:58,720 but no logs. 366 00:12:58,720 --> 00:13:00,800 Nothing helpful, you're stuck. 367 00:13:00,800 --> 00:13:02,560 The old way of handling this is painful. 368 00:13:02,560 --> 00:13:05,440 You open the portal, you dig through the container app settings, 369 00:13:05,440 --> 00:13:06,880 you check the deployment logs, 370 00:13:06,880 --> 00:13:09,880 you write CLI commands to pull diagnostic information, 371 00:13:09,880 --> 00:13:12,600 you compare configuration files between environments, 372 00:13:12,600 --> 00:13:14,680 you manually trace through the startup sequence 373 00:13:14,680 --> 00:13:17,680 trying to figure out what's broken, this takes hours, 374 00:13:17,680 --> 00:13:19,640 and it assumes you already know what to look for. 375 00:13:19,640 --> 00:13:22,840 With Azure MCP server, the experience is completely different. 376 00:13:22,840 --> 00:13:25,040 You copy the error message, paste it into co-pilot, 377 00:13:25,040 --> 00:13:26,800 enable the Azure MCP tools 378 00:13:26,800 --> 00:13:28,400 and let the agent do the diagnosis. 379 00:13:28,400 --> 00:13:29,240 That's it. 380 00:13:29,240 --> 00:13:31,760 The agent reads your project files, it looks at your app settings. 381 00:13:31,760 --> 00:13:33,680 It checks the container app configuration. 382 00:13:33,680 --> 00:13:35,200 And it identifies the issue. 383 00:13:35,200 --> 00:13:37,320 Your Azure app configuration connection string 384 00:13:37,320 --> 00:13:38,760 is a placeholder value. 385 00:13:38,760 --> 00:13:40,160 It was never configured for production. 386 00:13:40,160 --> 00:13:41,760 When the container app tries to start, 387 00:13:41,760 --> 00:13:44,400 it can't connect to app config, throws an exception, 388 00:13:44,400 --> 00:13:45,720 and fails to start. 389 00:13:45,720 --> 00:13:47,200 The agent doesn't just find the problem. 390 00:13:47,200 --> 00:13:48,640 It proposes a fix. 391 00:13:48,640 --> 00:13:50,320 Create a real app configuration store. 392 00:13:50,320 --> 00:13:52,600 Set the environment variables in the container app. 393 00:13:52,600 --> 00:13:54,880 Update the code to read from environment variables 394 00:13:54,880 --> 00:13:57,160 instead of hard-coded placeholders redeploy. 395 00:13:57,160 --> 00:13:58,640 And it walks you through each step, 396 00:13:58,640 --> 00:14:00,840 asking for permission before making changes. 397 00:14:00,840 --> 00:14:02,320 Here's where it gets impressive. 398 00:14:02,320 --> 00:14:05,680 The agent orchestrates multiple Azure operations and sequence. 399 00:14:05,680 --> 00:14:07,400 It queries Azure resource graph 400 00:14:07,400 --> 00:14:09,800 to find the container apps related to your project. 401 00:14:09,800 --> 00:14:11,480 It identifies the resource group. 402 00:14:11,480 --> 00:14:14,480 It creates the app configuration store in that same group. 403 00:14:14,480 --> 00:14:17,920 It updates the container app with the new environment variable. 404 00:14:17,920 --> 00:14:19,640 It makes the code changes in your project. 405 00:14:19,640 --> 00:14:21,040 It kicks off the deployment. 406 00:14:21,040 --> 00:14:22,920 And then it verifies the endpoint is working. 407 00:14:22,920 --> 00:14:24,320 The whole thing takes minutes. 408 00:14:24,320 --> 00:14:26,480 The agent handled every Azure operation. 409 00:14:26,480 --> 00:14:29,400 Reading resources, creating new ones, updating configurations, 410 00:14:29,400 --> 00:14:31,600 triggering deployments, verifying results. 411 00:14:31,600 --> 00:14:33,360 You didn't write a single CLI command. 412 00:14:33,360 --> 00:14:34,680 You didn't navigate the portal. 413 00:14:34,680 --> 00:14:37,080 You just described the problem and approved the steps. 414 00:14:37,080 --> 00:14:39,600 What this means for you is that complex troubleshooting 415 00:14:39,600 --> 00:14:42,240 becomes a conversation instead of a manual process. 416 00:14:42,240 --> 00:14:44,320 You don't need to be an Azure expert to diagnose 417 00:14:44,320 --> 00:14:45,280 and fix issues. 418 00:14:45,280 --> 00:14:46,800 The agent brings the expertise. 419 00:14:46,800 --> 00:14:48,640 You bring the context and the approval. 420 00:14:48,640 --> 00:14:50,560 And you can try this yourself with a test app. 421 00:14:50,560 --> 00:14:51,840 You don't need special permissions. 422 00:14:51,840 --> 00:14:53,200 You don't need to be a developer. 423 00:14:53,200 --> 00:14:54,720 You just need an Azure subscription 424 00:14:54,720 --> 00:14:57,560 and an MCP compatible AI assistant. 425 00:14:57,560 --> 00:14:59,720 How it all connects, the big picture. 426 00:14:59,720 --> 00:15:01,280 So now you understand the parts. 427 00:15:01,280 --> 00:15:02,880 MCP is the standard protocol. 428 00:15:02,880 --> 00:15:05,800 Azure MCP server is Microsoft's implementation. 429 00:15:05,800 --> 00:15:08,120 And Azure services are the tools it connects to. 430 00:15:08,120 --> 00:15:10,080 But the real value isn't any single piece. 431 00:15:10,080 --> 00:15:12,360 It's that they all work together through one protocol. 432 00:15:12,360 --> 00:15:13,640 What does that look like in practice? 433 00:15:13,640 --> 00:15:16,560 Imagine a single agent that can query Azure Monitor logs, 434 00:15:16,560 --> 00:15:19,000 check app configuration settings, create new resources 435 00:15:19,000 --> 00:15:21,600 and deploy updated code all in one session. 436 00:15:21,600 --> 00:15:23,280 That's not four separate integrations. 437 00:15:23,280 --> 00:15:26,280 That's one MCP server doing four things in sequence 438 00:15:26,280 --> 00:15:28,360 with context carried through each step. 439 00:15:28,360 --> 00:15:31,040 The output of one tool becomes the input for the next. 440 00:15:31,040 --> 00:15:33,280 The session state keeps everything connected. 441 00:15:33,280 --> 00:15:34,960 This creates interaction effects. 442 00:15:34,960 --> 00:15:36,720 You don't get with isolated tools. 443 00:15:36,720 --> 00:15:39,480 The agent reads logs, spots a configuration problem, 444 00:15:39,480 --> 00:15:41,000 builds a fix and deploys it. 445 00:15:41,000 --> 00:15:43,000 Without you having to manually copy information 446 00:15:43,000 --> 00:15:43,920 between steps. 447 00:15:43,920 --> 00:15:45,480 The network effect kicks into. 448 00:15:45,480 --> 00:15:47,240 As more MCP servers are built, 449 00:15:47,240 --> 00:15:50,280 every AI assistant gets access to more capabilities. 450 00:15:50,280 --> 00:15:52,080 Microsoft already has MCP servers 451 00:15:52,080 --> 00:15:54,640 for Azure DevOps, playwright and document processing. 452 00:15:54,640 --> 00:15:56,040 Third parties are building them as well. 453 00:15:56,040 --> 00:15:59,080 Each new server adds tools that any MCP compatible agent 454 00:15:59,080 --> 00:15:59,880 can use. 455 00:15:59,880 --> 00:16:02,880 The ecosystem grows more valuable with every server added. 456 00:16:02,880 --> 00:16:04,560 Now let's look at where this is headed. 457 00:16:04,560 --> 00:16:08,480 As your MCP server is already built into Visual Studio 2026. 458 00:16:08,480 --> 00:16:11,520 The Azure DevOps MCP server is generally available. 459 00:16:11,520 --> 00:16:13,880 Azure API management now supports MCP. 460 00:16:13,880 --> 00:16:16,560 So enterprises can govern and secure their MCP tools 461 00:16:16,560 --> 00:16:18,280 the same way they govern their APIs. 462 00:16:18,280 --> 00:16:20,640 And Microsoft announced agent orchestrator 463 00:16:20,640 --> 00:16:22,240 at Build 2026. 464 00:16:22,240 --> 00:16:24,360 A managed service that handles load balancing, 465 00:16:24,360 --> 00:16:26,080 health monitoring and cost attribution 466 00:16:26,080 --> 00:16:27,720 across thousands of agents. 467 00:16:27,720 --> 00:16:30,320 Early adopters saw a 30 to 45% reduction 468 00:16:30,320 --> 00:16:33,280 in per agent compute costs through intelligent model routing. 469 00:16:33,280 --> 00:16:34,680 The big picture is clear. 470 00:16:34,680 --> 00:16:37,920 MCP is becoming the universal integration layer for AI. 471 00:16:37,920 --> 00:16:41,040 And Azure MCP server is the on ramp for Azure users. 472 00:16:41,040 --> 00:16:44,360 Organizations that adopt MCP now will be able to plug into 473 00:16:44,360 --> 00:16:47,240 any AI platform without rebuilding their integrations. 474 00:16:47,240 --> 00:16:49,440 Your tools, your data, your security model, 475 00:16:49,440 --> 00:16:52,480 all accessible through one protocol from any AI assistant 476 00:16:52,480 --> 00:16:53,640 that speaks it. 477 00:16:53,640 --> 00:16:55,440 That's the knowledge nugget for today. 478 00:16:55,440 --> 00:16:58,720 MCP gives AI agents a standard way to talk to tools. 479 00:16:58,720 --> 00:17:00,640 And Azure MCP server gives them access 480 00:17:00,640 --> 00:17:02,520 to your entire Azure environment. 481 00:17:02,520 --> 00:17:04,000 If you're building with AI on Azure, 482 00:17:04,000 --> 00:17:05,520 start exploring MCP today. 483 00:17:05,520 --> 00:17:07,520 Subscribe on your favorite podcast platform 484 00:17:07,520 --> 00:17:09,520 and share this with someone starting their journey. 485 00:17:09,520 --> 00:17:11,120 Drop a comment if something clicked. 486 00:17:11,120 --> 00:17:12,320 I read everyone. 487 00:17:12,320 --> 00:17:13,840 I'm Moco Peters and this has been 488 00:17:13,840 --> 00:17:15,800 another episode of Microsoft Knowledge Nuggets.