65% of MCP Tools Now Take Actions. 16 Months Ago It Was 27%.

A new preprint analyzed 177,436 MCP tools deployed between November 2024 and February 2026. The headline number: action tools — tools that directly modify external environments — went from 27% to 65% of total MCP tool usage in 16 months. That's a 2.4x increase in the share of tools that do things to the world, not just read from it. The study is "How are AI Agents Used? Evidence From 177,000 MCP Tools" by Merlin Stein. It's the first large-scale empirical analysis of what agents are actually doing through MCP — not what we think they're doing, not what demos show, but what 177,000 deployed tools reveal about real-world agent behavior. From Reading to Writing The paper classifies MCP tools into three categories: perception (read data), reasoning (analyze data), and action (modify external environments). In late 2024, agents were mostly reading. Perception tools dominated. The typical MCP server was a data access layer — agents queried databases, read files, fetched API responses. By Feb