A Large-Scale Dataset of MCP Implementations on GitHub
This paper introduces the first large-scale, evidence-based dataset of 2,297 validated Model Context Protocol (MCP) implementations from GitHub, created through a rigorous hybrid verification pipeline to establish a foundational benchmark for analyzing real-world MCP ecosystems, development trends, and architectural patterns.
Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine the world of AI as a giant, chatty robot that lives inside a computer. For a long time, this robot could only talk about what it learned in its training data, like a student who only knows what's in their textbook. But recently, a new rulebook called the Model Context Protocol (MCP) arrived. Think of MCP as a universal translator and a set of standard power outlets that let the robot plug into real-world tools—like weather apps, search engines, or code editors—so it can actually do things instead of just talking about them.
But here's the problem: nobody knew exactly how many people were actually building these "power outlets" or what they looked like under the hood. There were rumors and lists, but no solid map of the real code.
That's where this paper comes in. The authors acted like digital treasure hunters, diving into GitHub (a giant library where developers store their code) to build the first massive, verified map of these MCP projects.
The Great Dig
The team didn't just guess; they built a super-smart robot pipeline to hunt for clues. They started with 3,238 potential projects that mentioned "MCP." But not every project that says "I'm an MCP project" actually is one. Some were just empty shells, others were just tutorials, and some were using the name "MCP" for something totally different (like a game mod for Minecraft, not an AI tool).
To get to the truth, they ran a multi-stage check-up:
- The Resume Check: Did the project have recent updates? If it hadn't moved in nine months, it was out.
- The Code Scan: They looked for specific "fingerprint" files (like
package.jsonorpyproject.toml) that prove the project is actually trying to connect to the MCP system. - The Structure Check: They looked for specific folders and entry points that show the code is ready to run, not just sitting there as a concept.
After all this filtering, they found 2,297 projects that were the real deal. They were so careful that when they manually double-checked a random sample, they found the process was 83% accurate. They even spotted 90 projects that looked like real tools but were actually just "show-and-tell" templates with no working parts, and they kicked those out of the final list.
What Are These Projects Doing?
Once they had their list of 2,297 verified projects, they sorted them into three main roles, like characters in a play:
- The Clients: These are the "requesters." They are like the robot's hands, reaching out to ask for tools. (Think of a chat interface asking for the weather).
- The Servers: These are the "providers." They hold the tools and hand them over when asked.
- The Gateways: These are the "middlemen" or translators. They help different systems talk to each other, bridging gaps between the AI and outside services.
The data suggests that Python and TypeScript are the superstar languages for building these tools. They dominate the scene for both the requesters (Clients) and the providers (Servers). Interestingly, the "middlemen" (Gateways) also like to use Go, a language known for speed and reliability.
How Are They Built?
The paper suggests that most of these projects aren't massive, corporate monoliths. Instead, they look like small, agile teams. The typical project has about 4 unique contributors and 70 commits (updates). The developers seem to prefer making small, frequent tweaks rather than huge, scary updates all at once.
What This Means
This paper doesn't claim to have solved every mystery about AI tools. Instead, it provides a solid, evidence-based foundation. It suggests that while the ecosystem is growing fast, it's currently dominated by a few key languages and follows a specific pattern of small, iterative development. By giving researchers a clean, verified list of 2,297 real projects, this work sets the stage for future studies to understand how these AI connections are really evolving, rather than just guessing based on marketing hype.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.