AI Client Test Reveals Server Communication Failures and Token Discrepancies

A recent test involving 90 trials of the Model Context Protocol (MCP) revealed that one client failed to reach the server for most of its calls, mimicking a model that performed poorly. While 87 out of 90 trials completed successfully, the failures highlighted differences in how Claude Code and Gemini CLI handled tasks and negotiated protocol revisions. Notably, one GitHub task resulted in a significant token count discrepancy between the two clients, with Claude Code receiving 1,561 tokens for a response while Gemini CLI received only 161. AI

IMPACT Highlights potential issues in AI agent communication protocols and token usage, suggesting areas for improvement in reliability and efficiency.

RANK_REASON The item details a technical experiment and its findings regarding AI client-server communication and performance, fitting the research category. [lever_c_demoted from research: ic=1 ai=1.0]

Read on dev.to — MCP tag →

• Claude Code
• Claude Haiku 4.5
• Claude Sonnet-5
• file system
• Gemini 2.5-Flash
• Gemini CLI
• GitHub
• MCP
• @modelcontextprotocol/server-filesystem
• playwright
• Playwright MCP

• product
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →