Izzo: Your AI assistant just asked to update your CRM, check inventory, and send a Slack message. Your IT team is having nightmares.
Izzo: You're listening to Exploring Next, episode two-forty-six. I'm Izzo, and with me is Boone. Today we're talking MCP — Model Context Protocol — and why every IT team should care about this right now.
Boone: Because we're about to hit this wall where AI agents can do incredible reasoning but can't actually touch any of our business systems in a standardized way.
Izzo: Exactly. It's like having a brilliant intern who speaks twelve languages but doesn't know how to use any of our tools. So Boone, what exactly is MCP solving here?
Boone: Think of it as a universal translator between AI models and enterprise systems. Instead of every AI tool building custom integrations to your CRM, database, file system — MCP creates a standard protocol.
Izzo: Okay but break that down for me. What does this protocol actually look like under the hood?
Boone: It's built on JSON-RPC, which gives you bidirectional communication. So your AI client — could be Claude, could be a custom agent — connects to MCP servers that wrap your business systems.
Izzo: And these servers are doing what exactly?
Boone: Three main things. They expose tools — like 'update customer record' or 'query inventory.' They provide resources — access to files, databases, APIs. And they handle prompts — context about how to use all this stuff.
Izzo: So instead of every AI vendor writing their own Salesforce connector, they just need to speak MCP and connect to your Salesforce MCP server.
Boone: Exactly. And the security model is built in. The MCP server handles authentication, permissions, rate limiting — all the stuff you need for production.
Izzo: This feels like API management all over again, but for AI. Who's actually shipping this? Because I'm seeing a lot of protocols that sound great in demos but die in implementation.
Boone: Anthropic open-sourced the spec and they're already using it in Claude Desktop. You can connect to GitHub, file systems, databases — it's not vaporware.
Izzo: That's smart positioning by Anthropic. Get the ecosystem building MCP servers while Claude becomes the reference client.
Boone: And the implementation is surprisingly clean. Each MCP server exposes a schema that describes its capabilities. The AI client can discover what tools are available and how to use them dynamically.
Izzo: Wait, so the AI can basically ask 'what can you do?' and get a structured response?
Boone: Right. It's like OpenAPI specs but designed specifically for AI consumption. The server says 'I have a create_ticket tool that takes title, description, and priority parameters.'
Izzo: That's actually brilliant for adoption. IT teams can wrap existing systems without changing them, and AI tools get instant access. What's the catch?
Boone: Orchestration gets complex fast. Once you have agents calling multiple MCP servers, managing state and error handling across systems becomes your new problem.
Izzo: Yeah, I can see that. It's like microservices — great until you need to coordinate twelve different services for one business process.
Boone: Exactly. And you're probably going to need some kind of workflow engine or orchestrator sitting above MCP to handle the complex multi-step operations.
Izzo: Which brings us back to the original challenge — most enterprises aren't architected for this kind of AI-driven automation. This is infrastructure work.
Boone: But that's also the opportunity. Companies that get their MCP servers built and their orchestration layer right are going to have a huge advantage when AI agents actually go mainstream.
Izzo: I'm giving MCP itself an A-minus. Clean spec, real implementation, solves a genuine problem. The orchestration layer above it is where things get interesting.
Boone: Agreed. And honestly, I'm already adding an MCP server for our internal tools to the weekend project list. The barrier to entry is really low.
Izzo: Alright, for people who want to get hands-on with this — Boone, what should they actually go build?
Boone: Start simple. Clone the MCP Python SDK from GitHub and build a server that wraps one internal API. Test it with Claude Desktop to see the magic happen.
Izzo: And if you want to go deeper, look at the existing MCP servers — there's ones for PostgreSQL, GitHub, file systems. Study how they handle authentication and error cases. Third thing — if you're in enterprise IT, start identifying which systems would benefit most from AI agent access. CRM, ticketing, monitoring tools — that's your MCP server roadmap. MCP isn't just a protocol — it's infrastructure for the AI-first enterprise that's coming whether we're ready or not. Time to st