agentwhy

agentwhy is made by Karol Kozer. It reads the conversation records that Claude Code, Codex and GitHub Copilot already save on your computer and reports which private files (such as .env files and keys) an AI agent or its helpers opened, the task that led to it, and which key to change. It runs locally via npx @agentwhy/cli, and the repository lists a v0.4.2 release and an Apache 2.0 license.

What It Is

agentwhy is a local security-visibility tool for people who build with AI coding agents. You run one command in your project folder, and a page opens on your computer listing conversations, the private files reached, who reached them, and what to fix. The project describes itself as pre-MVP. It processes records by fixed rules, with no AI model, account, or internet connection, and the README says secret values are replaced before a report is built.

How the report works

The dashboard shows a week of conversations, each marked by whether a private file was opened. Opening a conversation shows a to-do list, a step-by-step story of what happened, a diagram, a by-folder view, and a full event record. It also traces helper agents: which helper was given what task, which files it touched, and what it passed back to the main agent. For each private file you can choose to block it, so the agent is stopped, or track it, so the agent may read it and you are told.

Optional protection

init can add rules and hooks to Claude Code, Codex, and (when enabled) Copilot that refuse obvious reads of private files. It shows exactly what it will write and asks first, and init --remove undoes it. The documentation states this is not a sandbox: paths built at runtime, scripts, or commands handed to another shell can still get through, which is why the report reads what actually happened. The README also notes that the report can miss file access the transcript does not record, such as reads inside scripts or via connected MCP tools.

Platform and agent support

It requires Node 22 or later and was made and tested on a Mac; Windows is described as still being tested. It supports Claude Code, Codex, and GitHub Copilot conversations, but not Cursor's AI, Windsurf, or chats on claude.ai.