MCP Servers: When They're Worth It and When You're Adding Middleware for Nothing

Do you actually need MCP? Probably not. Model Context Protocol is a few lines of code to implement. But the moment you deploy it in the cloud, you're not dealing with a protocol anymore. You're dealing with infrastructure. CloudFront, WAF, ALB, VPC, ECS Fargate, Cognito, Secrets Manager, four CDK stacks just to wire it together. That's the recommended AWS architecture. MCP solves a real problem: giving agents a standard interface to tools they don't own. But standardization comes at a price. Blocking responses, serialization overhead, no push pattern, Base64-encoded binary data, type flattening. These aren't bugs. They're trade-offs. And most teams adopting MCP don't actually need what it provides. In this video we break down exactly when MCP makes sense, when it doesn't, and the three questions you should ask before committing to it. Plus a concrete side-by-side comparison showing what MCP adds to a simple database query you could handle with a direct function call. If your agent talks to systems you don't own, MCP is the right call. If it talks to your own infrastructure, you might be adding middleware for a problem you don't have. If you want to implement this architecture in your business, let's talk: https://www.linkedin.com/in/tiagomrsantos/ https://www.linkedin.com/company/tensorops/ AWS GitHub repo: https://github.com/aws-solutions-library-samples/guidance-for-deploying-model-context-protocol-servers-on-aws