The Boring Infrastructure Beneath MCP Is the Real Story

Something shifted this week. Not another model release, but the infrastructure underneath. AWS shipped an MCP server. Unity got one. So did a code search tool built on a vector database company's stack. None of that is glamorous. All of it matters more than whatever benchmark chart did the rounds on Monday.

Here's the thing people miss about the Model Context Protocol. It was never really about Claude talking to more tools. It was about who builds the bridge and who has to maintain it forever afterwards. When AWS Labs puts its name on an MCP server, Amazon is deciding this protocol is worth institutional weight, not a side project some engineer built over a weekend and abandoned. The same goes for a game engine vendor letting an LLM reach into the Unity Editor. A company has decided its product needs an AI-shaped front door. That's a real signal, not a toy demo.

The pattern worth watching is who's building these and why. AWS is in. So is Prefect, so is a search infrastructure company, so is an open source coding toolkit that grew a community without any of that backing at all. Anthropic barely features. MCP won by becoming boring enough that unrelated companies build on it without asking permission first. Nobody announces it. It just starts turning up in places you didn't expect.

The number that should stop you is the star count on some of these. A toolkit built as a coding IDE substitute, from a small studio, sitting past 27,000 stars. Thousands of developers decided a tool solved something painful enough to bother starring it, and that's a much higher bar than clicking a link in a newsletter.

Traction like that is a filter, not a discovery. A repo with no stars and no commits tells you nothing. A repo with tens of thousands does. This week gave us five that cleared it, and every single one came from a different corner of the stack: cloud infrastructure, game development, code search, developer tooling, and the raw framework underneath all of it.

AWS Puts Its Name Behind an MCP Server AWS Labs MCP (GitHub) is exactly what it sounds like: an official, open source MCP server implementation from Amazon, sitting at 9,523 stars. It lets Claude and other MCP clients talk to AWS services directly, which matters more than the star count suggests. AWS doesn't put its name on side projects. When a company this size ships infrastructure for a protocol, it's telling developers the protocol is safe to build a business on. That's a different signal from a startup's weekend release. Treat this as the moment MCP stopped being an Anthropic bet and started being shared plumbing.

A Coding IDE Built Entirely on MCP 27,000 stars is a serious number for anything, and Serena (GitHub) has earned it. It's a Python-based MCP toolkit that does semantic code retrieval and editing, positioned to function like an IDE rather than a bolt-on. Developers using Claude for coding work have been asking for exactly this: something that understands a codebase's structure, not just its text. A community voted with stars because the tool solved a real problem.

Your Whole Codebase, Not Just the Open File Context is the thing that breaks Claude Code on large projects, and Claude Context (GitHub) from Zilliz goes straight at it. Built in TypeScript, it lets you use an entire codebase as context for a coding agent rather than whatever fits in a single prompt window. Over 12,000 stars back it up. The pain point is real. Anyone who's watched Claude lose track of a function defined three files away knows exactly why this exists. Worth a look if that's been your ceiling.

Prefect's Python Framework for Building MCP Servers Nearly 27,000 stars, and it comes from an established team rather than an unknown. FastMCP (GitHub), built by PrefectHQ, gives Python developers a proper framework for building MCP servers and clients rather than hand-rolling the protocol themselves. This is the least visible item on this list and probably the most useful one. Every server above needed something like this underneath it. If you're building your own MCP integration rather than just using someone else's, start here.

Where do you think MCP goes next now that AWS and Unity are building on it too?

Tell me in the comments.

Forward this to one person who should be using AI better than they are. Reply with what you built, tried or broke this week. I read every one.

Gareth, founder of The Anthropic Stack (theanthropicstack.com)