CurseBench

Overview

CurseBench scans local coding-agent chat stores and counts abusive terms without sending chat text anywhere. It compares harnesses and models by total uses, message rate, and affected sessions, then renders a terminal dashboard or writes a machine-readable JSON report.

Highlights

• Reads Claude Code, Codex, OpenCode, Antigravity, Cursor, and Grok Build session stores
• Every adapter is read only, and message excerpts stay hidden unless explicitly requested
• Ignores matches inside code and excludes common technical phrases to reduce false positives
• Reports parser health so missing stores and incompatible formats are visible

Stack

Python 3.11+

CLI, local store adapters, and report generation

Standard library

Dependency-free runtime

JSON

Machine-readable benchmark reports