An Empirical Study of Model Context Protocol Applications

2026-07-28

AI Summary

This study addresses the lack of standardized practices in Model Context Protocol (MCP) regarding configuration, communication, and human oversight, noting that prior research has predominantly focused on server-side implementations while neglecting application-level usage. To bridge this gap, we introduce MCPAppTax, the first taxonomy for MCP applications, and conduct a large-scale empirical analysis of 1,723 MCP applications on GitHub, leveraging large language model–assisted annotation and static code analysis. Our findings reveal both convergent practices—such as 85.2% adopting file-based configuration and 81.1% using official SDKs—and divergent ones, notably the absence of standardized naming conventions for configuration parameters. Furthermore, while human supervision mechanisms are prevalent (90.8% log interactions and 77.2% offer start/stop controls), only 37.2% implement blocking-style human approval, highlighting significant variability in safety-critical oversight.

Abstract

The Model Context Protocol (MCP) standardizes how large language model applications communicate with external tools, but leaves the application side unspecified: unlike traditional dependencies resolved through package managers, developers integrating MCP servers face no conventions for configuration, communication, or human oversight. This ecosystem is also under-researched, with existing work focused on servers rather than the applications consuming them. We conduct a large-scale study of 1,723 MCPApps mined from GitHub. We first derive MCPAppTax from a representative sample, then use an LLM-assisted pipeline to apply it across the full dataset, characterizing server integration across configuration, SDK use, and human-in-the-loop mechanisms. Our results show that the ecosystem has converged on some practices but not others: most MCPApps configure servers using files (85.2%) and use an official SDK (81.1%) to communicate with servers, yet no naming convention has emerged for configuration files. Human oversight diverges most, logging (90.8%) and enable/disable controls (77.2%) are common, but only 37.2% gate tool execution behind a blocking approval step, leaving the LLM able to invoke any enabled tool unconditionally in most MCPApps.

Problem

• Model Context Protocol
• large language model applications
• tool integration
• human oversight
• configuration conventions

Innovation

• Model Context Protocol
• LLM-assisted analysis
• empirical study
• human-in-the-loop
• tool integration