confessor

local forensic report of what an AI coding agent accessed

TLDR

Scan local Claude Code session logs and open an HTML report

$ npx confessor

Scan a chat export zip or directory

$ npx confessor [~/Downloads/claude-export.zip]```

Write report to a **path** and skip opening the browser

Emit JSON and fail CI on high-severity findings

$ npx confessor --json --fail-on [high]

# SYNOPSISconfessor [options] [export-path]# DESCRIPTIONconfessor reconstructs what an AI coding agent did from logs already on disk—primarily Claude Code session JSONL under ~/.claude/projects—and builds an offline HTML report. It lists files read or written, secrets that entered tool results, network or MCP sinks, and exposure paths: sensitive reads followed by outbound activity in the same session.The same engine can scan ChatGPT, Claude.ai, and Gemini Takeout exports for pasted secrets and sensitive topics. Detection uses rule patterns (no ML). Secrets are redacted before storage. The tool makes zero network calls itself and has zero runtime dependencies beyond Node.# PARAMETERS--out file

Report output path (default HTML name)

--no-open

Do not open the report in a browser

--json

Machine-readable output

--quiet

Reduce logging

--fail-on critical|high|medium

Exit non-zero if findings meet or exceed severity (for CI)

# CAVEATSExposure paths are investigative leads, not proof of theft. Agent reconstruction currently targets Claude Code log format; other agents may not be fully supported. Retrospective only—does not intercept live sessions.# SEE ALSO[claude](/man/claude)(1)# RESOURCES

$ Source code

$ Homepage