Abstract
The Model Context Protocol (MCP) has recently emerged as a standardized interface for connecting language models with external tools and data. As the ecosystem rapidly expands, the lack of a structured, comprehensive view of existing MCP artifacts presents challenges for research. To bridge this gap, we introduce MCPCorpus, a large-scale dataset containing around 14K MCP servers and 300 MCP clients. Each artifact is annotated with 20+ normalized attributes capturing its identity, interface configuration, GitHub activity, and metadata. MCPCorpus provides a reproducible snapshot of the real-world MCP ecosystem, enabling studies of adoption trends, ecosystem health, and implementation diversity. To keep pace with the rapid evolution of the MCP ecosystem, we provide utility tools for automated data synchronization, normalization, and inspection. Furthermore, to support efficient exploration and exploitation, we release a lightweight web-based search interface. MCPCorpus is publicly available at: https://github.com/Snakinya/MCPCorpus.
Introduction
As Large Language Models (LLMs) transition from simple chat interfaces to "agentic" workflows, they require a standardized method to interact with the physical and digital world—reading local files, querying databases, or triggering API calls. The Model Context Protocol (MCP) has emerged as the leading interface for this "brain-to-tool" connection.
However, the MCP ecosystem is currently in a state of rapid, decentralized expansion. This "wild west" environment lacks a centralized, structured repository, making it difficult for researchers to audit security vulnerabilities (such as command injection) or for developers to ensure interoperability across different LLM platforms. Previous research has been limited to small, hand-picked samples, which fail to capture the full diversity of the ecosystem.
The MCPCorpus Dataset
To address these challenges, researchers from the National University of Singapore have introduced MCPCorpus, the first large-scale, multi-source dataset designed to map the MCP landscape.
Scale and Scope
- Volume: The dataset includes approximately 13,875 MCP servers and 300 MCP clients.
- Richness: Each entry contains up to 26 structured fields, combining registry information with technical and social signals derived from GitHub.
- Format: Data is distributed in newline-delimited JSON (JSONL) for easy programmatic access and integration into research pipelines.
Methodology
The authors developed a five-stage automated pipeline to construct and maintain the corpus, ensuring it remains a living resource:
- Registry Crawling: The pipeline systematically scrapes
MCP.so, the largest community registry, to extract basic metadata, tool schemas, and deployment interfaces. - GitHub Metadata Enrichment: For artifacts with source code on GitHub, the pipeline queries the GitHub REST API to pull signals such as star counts, forks, contributor activity, and primary programming languages.
- Cleaning and Deduplication: To ensure data quality, unreachable repositories are removed, and canonicalized URLs are used to filter out forks or mirrors.
- Metadata Normalization: Disparate data points are mapped into a unified schema with over 20 attributes, including license types, server commands, and Dockerfile presence.
- Classification and Indexing: Artifacts are categorized by type (server vs. client), application domain (e.g., AI, Cloud, Data), and language stack.
Key Insights into the MCP Ecosystem
Analysis of MCPCorpus reveals several critical trends regarding how tools are being built and shared:
The "Long Tail" of Popularity
The ecosystem is highly top-heavy. While a few flagship projects, such as Microsoft's playwright-mcp, boast over 11,000 stars, more than 10,000 projects have fewer than 10 stars. This indicates a massive wave of experimental, "long-tail" development by individual hobbyists alongside a few consolidated industry tools.
Language Hegemony
Python and TypeScript/JavaScript are the undisputed dominant languages in the MCP world. This consistency holds true across both experimental projects and well-maintained, popular tools, suggesting a standardized developer experience is forming around these stacks.
Security and Attack Surface
By capturing fields like server_command and server_config, MCPCorpus exposes the "attack surface" of these tools. Researchers can now perform bulk analysis to identify unsafe endpoint exposures or missing authentication patterns across thousands of implementations. Importantly, this metadata is extracted via static inspection, allowing for comprehensive auditing without the security risks of executing untrusted third-party code.
Utility and Accessibility
To ensure the dataset remains relevant as the ecosystem grows, the authors have released utility tools for automated synchronization and normalization. Additionally, a lightweight web-based search interface has been provided to allow developers and researchers to explore the dataset efficiently.
Limitations and Future Work
While MCPCorpus provides a vital foundation, it currently relies on static metadata. Consequently, it cannot confirm the runtime behavior or functional efficacy of the tools listed. Future work aims to incorporate sandboxed runtime testing to validate tool performance and security. Furthermore, the authors plan to expand the dataset's reach by indexing artifacts from other package managers like npm and PyPI, and by refining classification to distinguish between hobbyist and corporate-backed tools.
Conclusion
MCPCorpus provides the first comprehensive map of the Model Context Protocol ecosystem. By unifying data from centralized registries and decentralized repositories, it enables a new wave of research into LLM-to-tool integrations, security auditing, and the evolution of agentic AI.
Summary updated