MCPZoo: 大规模可运行模型上下文协议服务器数据集（面向AI代理）

Jul 16, 2026

Upload by Anda

Model Context Protocol (MCP) allows AI agents to connect to external tools, yet empirical study has been limited by the absence of large, ready-to-use datasets. MCPZoo introduces a large-scale collection of MCP servers from multiple public sources, totaling 129,059 servers with 56,053 distinct entries. It further includes 16,356 runnable and interactable server instances, validated via deployment-oriented checks. The dataset supplies unified metadata and access interfaces for systematic, reproducible experimentation and security analysis.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages3

Reading Time8 min

Introduction

Build Infrastructure

Construct MCP Zoo

Make the Zoo Alive

MCPZoo: A Large-Scale Dataset of Runnable Model Context Protocol Servers for AI Agent Mengying Wu†, Pei Chen†, Geng Hong†B, Baichao An†, Jinsong Chen†, Binwang Wan†, Xudong Pan†‡, Jiarun Dai†, and Min Yang†B †Fudan University, ‡Shanghai Innovation Institute {wumy21, jschen23, [bwwan25}@m.fudan.edu.cn](bwwan25}@m.fudan.edu.cn), {peichen19, ghong, bcan20, xdpan, jrdai, [m_yang}@fudan.edu.cn](m_yang}@fudan.edu.cn) BCo-corresponding authors arXiv :2512 . 15144v3 [ cs .CR] 26 Dec 2025 Abstract—Model Context Protocol (MCP) enables agents to interact with external tools, yet empirical research on MCPis hindered by the lack of large-scale, accessible datasets. We present MCPZoo, the largest and most comprehensive dataset of MCP servers collected from multiple public sources, comprising 129,059 servers (56,053 distinct). MCPZoo includes 16,356 server instances that have been deployed and verified as runnable and interactable, supporting realistic experimentation beyond static analysis. The dataset provides unified metadata and access interfaces, enabling systematic exploration and interaction without manual deployment effort. MCPZoo is released as an open and accessible resource to support research on MCP-based systems and security analysis. Index Terms—Model Context Protocol, MCP, Agents, Agent Security, Security Measurement 1. Introduction The software ecosystem is witnessing a paradigm shift driven by Large Language Models (LLMs) and autonomous Agents. Just as mobile applications revolutionized digital interaction over the last decade, "Tools"—interfaces that allow Agents to perceive and manipulate the world—are becoming the fundamental building blocks of the AI era. The Model Context Protocol (MCP) has rapidly emerged asthe de facto standard for connecting these AI models with diverse data sources and environments. However, conducting large-scale empirical research on this growing ecosystem presents significant challenges. While recent studies have begun to explore the security risks and robustness of MCP integrations (e.g., prompt injection vulnerabilities in tool use), these efforts are often constrained by the difficulty of setting up testing environments. Unlike mobile apps, which are packaged as standalone binaries, MCP servers are diverse software projects with complex, often conflicting dependencies. Consequently, existing research is frequently limited to small, manually curated datasets, leading to experiments that are difficult to reproduce or generalize. Inspired by AndroZoo [1], which democratized Android research by collecting millions of apps for the community, we present MCPZoo. To the best of our knowledge, MCPZoo is the largest collection of Model Context Protocol servers to date. While AndroZoo solved the problem of access, MCPZoo aims to solve the problem of runnability. By providing a standardized, unrestricted, and scalable dataset of runnable agents, we aim to enable the research community to move beyond small-scale testing and conduct comprehensive, reproducible analyses on the security and capability of the AI agent ecosystem. Contribution. MCPZoo makes the following contributions: • The largest and most comprehensive dataset of MCP servers currently available, with 129,059 MCP servers. • The first dataset to include 16,356 runnable and interactable server instances. • Released as an open and accessible dataset to support community-driven MCP analysis. 2. Build Infrastructure 2.1. Construct MCP Zoo Sources. We collected MCP server projects from eight public sources: MCP Store [2], MCP World [3], MCP Market [4], MCP Servers Repository [5], AIbase MCP [6], Pulse MCP [7], [MCP.so](MCP.so) [8], and Smithery [9]. Each source provides an independently curated directory of MCP servers, reflecting different community and platform perspectives. Data Fields. Both metadata and source code for each MCP server are saved. We extract a unified set of metadata fields, including th

FAQs

What problem does MCPZoo address in MCP research?

It addresses the lack of large-scale, accessible MCP server datasets, which makes empirical research difficult to set up and reproduce.

How many MCP servers and runnable instances are included in MCPZoo?

MCPZoo contains 129,059 MCP servers (56,053 distinct) and 16,356 server instances that are deployed and verified as runnable and interactable.

How does MCPZoo make servers usable at scale?

It builds executable Docker images from server source code using an automated build agent, then performs standardized liveness checks using both execution-capability validation and basic protocol-level interactions.

Jul 16, 2026

Upload by Anda