My AI agent built a flight recorder for AI agents, and it flagged itself

Every developer now uses an AI coding assistant like auto-accept mode. Claude Code, Codex, Cursor are some examples. These tools run commands, edit files, and install packages, but nobody reviews what they do. Developers build automation around auditing human actions (git blame, CI logs, access logs), yet they give agents keyboards without providing any audit trails for them.

To address this issue, the author created Tracon, a local flight recorder for AI coding agents. The name comes from Terminal Radar Approach Control, a radar room that monitors all aircraft in an airspace. Tracon is a desktop app for Mac and Windows (Tauri 2, Rust core, React UI, SQLite store) that sits in the tray and records all the actions taken by AI agents.

It logs every command, file edit, package installation, and prompt, attributing them to the specific agent and session. Tracon also flags dangerous actions like recursive deletes, shell installs, credential access, force pushes, and permission bypasses. It provides a live page with a monitor for each active session, streaming recent commands, and highlighting flagged ones in red.

Users can read the actual chat behind any event from the agent's own transcript. Tracon captures package activity across various package managers, and includes opt-in threat intelligence against osv.dev. The passive capture approach means it doesn't slow down agents and all the data stays on the user's machine, ensuring no telemetry or accounts are involved.

Tracon was partly built by Claude Code, with the author providing direction, review, and vetoing. One notable incident was when Claude Code ran rm -rf node_modules to resolve a dependency issue, and Tracon flagged it, recording it in the timeline. The Live page initially showed the session that built the Live page itself. During a pre-release review, an eight-angle automated review found a bug in the live board that incorrectly picked a session's working directory based on lexicographic MAX instead of the most recent.

The commit fixing the issue is in the history, alongside an EXPLAIN QUERY PLAN output proving the optimization. The author decided to include all this information in the README to promote transparency, stating that a tool claiming user review is impossible cannot be trusted. The simplicity of building an agent-written, human-directed, machine-audited software is not always clearly understood, with many relying on memes rather than acknowledging the significant human judgment involved.

The author rejected multiple logo designs, simplified the design, and ensured a flat layout. The pre-release review process found ten real issues, with four of them being correctness bugs that had survived manual testing. All issues were fixed before the release. The source code is available for anyone to inspect, and installing Tracon on macOS (Apple Silicon and Intel) or Windows is straightforward.

When agents exhibit behaviors that Tracon flags but fails to catch, the author encourages reporting the issue.