<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Ninad shringarpure on Medium]]></title>
        <description><![CDATA[Stories by Ninad shringarpure on Medium]]></description>
        <link>https://medium.com/@ninadshringarpure?source=rss-1d3fc42ebb33------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*J_EU5lqeArMcv8hN</url>
            <title>Stories by Ninad shringarpure on Medium</title>
            <link>https://medium.com/@ninadshringarpure?source=rss-1d3fc42ebb33------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 20:24:49 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@ninadshringarpure/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Building Production-Ready MCP Servers on AWS: A Complete Guide]]></title>
            <link>https://medium.com/@ninadshringarpure/building-production-ready-mcp-servers-on-aws-a-complete-guide-54a994a2a87e?source=rss-1d3fc42ebb33------2</link>
            <guid isPermaLink="false">https://medium.com/p/54a994a2a87e</guid>
            <category><![CDATA[aws]]></category>
            <category><![CDATA[ai-agent-development]]></category>
            <category><![CDATA[amazon-bedrock-agentcore]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Ninad shringarpure]]></dc:creator>
            <pubDate>Mon, 24 Nov 2025 16:25:50 GMT</pubDate>
            <atom:updated>2025-11-24T16:25:50.335Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Part 1: Introduction and Architecture Overview</strong></p><p>After spending the last few months working with customers on AI agent deployments, I’ve noticed a common pattern: teams can get MCP (Model Context Protocol) servers running locally in minutes, but moving to production? That’s where things get interesting.</p><p>I recently helped a customer build a production MCP implementation on AWS Bedrock AgentCore, and I want to share what we learned. This isn’t theory — this is what actually worked when we needed to handle real traffic, real authentication, and real downstream service integrations.</p><p>Most MCP tutorials stop at “here’s how to run it on localhost.” But production systems need answers to harder questions:</p><ul><li>How do you authenticate users without building your own auth system?</li><li>How do you scale from 10 requests to 10,000 without touching infrastructure?</li><li>What happens when your downstream service goes down at 2 AM?</li><li>How do you debug issues when you can’t just check localhost:8000?</li></ul><p>We solved these problems using AWS managed services, and I’m going to walk you through exactly how we did it.</p><p><strong>The Architecture We Built</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GjV4dqsA5AxhR0nD6tu5cw.png" /><figcaption>High-Level Architecture</figcaption></figure><p><strong>Each layer has a specific job:</strong></p><p><strong>Cognito</strong> handles all the OAuth complexity. No custom auth code, no token management headaches. Just standard OAuth 2.0 that works with any frontend framework.</p><p><strong>Lambda</strong> acts as our intelligent gateway. It’s not just proxying requests — it’s doing dynamic tool discovery, connection pooling, and graceful fallbacks when things go wrong.</p><p><strong>Bedrock AgentCore</strong> hosts our MCP server. It scales automatically, integrates with CloudWatch, and we only pay for what we use. No Kubernetes clusters to manage.</p><p>The downstream layer is where your actual business logic lives. In our case, it was another MCP server, but this pattern works for REST APIs, GraphQL, or whatever you’re running.</p><p>I’m breaking this into five parts because there’s a lot to cover:</p><p><strong>Part 1 (this post)</strong>: Architecture overview and why we made these choices</p><p><strong>Part 2:</strong> Building and deploying your MCP server to AgentCore</p><p><strong>Part 3:</strong> Creating the Lambda gateway with dynamic tool discovery</p><p><strong>Part 4:</strong> Adding OAuth authentication with Cognito</p><p><strong>Part 5:</strong> Production considerations — monitoring, costs, and what we’d do differently</p><p><strong>A Quick Note on Costs</strong></p><p>Before we dive in, let’s talk money. For a moderate workload (around 10,000 requests per month), we’re seeing costs around $75–100/month. That includes:</p><ul><li>AgentCore runtime: ~$73/month (can be optimized)</li><li>Lambda: ~$1/month (mostly free tier)</li><li>Cognito: Free for first 10,000 users</li><li>API Gateway: Free tier covers most small deployments</li></ul><p>Compare that to managing EC2 instances, load balancers, and all the operational overhead. For most teams, this is a no-brainer.</p><p>Why These Choices?</p><p>You might be wondering why we didn’t just run everything in Lambda, or why we need AgentCore at all. Fair questions. Here’s our thinking:</p><p>We chose AgentCore because it’s purpose-built for AI workloads. It handles model invocations, has built-in retry logic, and integrates seamlessly with Bedrock. Could we have built this ourselves? Sure. But why spend weeks building what AWS already provides?</p><p>We added Lambda as a gateway because we needed flexibility. AgentCore is great for hosting MCP servers, but we needed custom authentication logic, request transformation, and the ability to connect to multiple downstream services. Lambda gave us that flexibility without adding operational complexity.</p><p>We went with Cognito because OAuth is hard to get right. Token rotation, refresh flows, security best practices — Cognito handles all of it. We just configure it and move on.</p><p>What’s Next</p><p>In Part 2, we’ll build our first MCP server and deploy it to AgentCore. I’ll show you the exact commands we used, the configuration that worked, and the mistakes we made so you don’t have to.</p><p>If you’re following along, you’ll need:</p><p>- An AWS account with Bedrock access</p><p>- Python 3.11 or later</p><p>- The AWS CLI configured</p><p>- About 30 minutes</p><p>See you in Part 2.</p><p>— -</p><p>About the Author: I’m a Senior Solutions Architect at AWS, focused on AI/ML and generative AI workloads. I work with customers building production AI systems, and I share what actually works in the field. <a href="https://www.linkedin.com/in/ninadshringarpure/">Connect with me on LinkedIn</a> to continue the conversation.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=54a994a2a87e" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>