Something weird is happening in software development right now. We're all writing code with AI tools like Claude Code, Codex CLI, and Cursor—but nobody really knows if we're doing it well, or even how to get better. Y Combinator's new experiment called Paxel wants to change that by giving developers a detailed breakdown of their own AI-assisted coding habits.

How It Works

Paxel analyzes the session transcripts stored locally on your machine from Claude Code, Codex CLI, and Cursor. You run a simple curl command in your terminal pointing at either a single repo or a parent directory containing multiple projects. The entire analysis runs inside a Docker container on your own machine—YC never sees your code, only transcript excerpts that get sent to Claude or GPT via their proxy for summarization. After 15 to 30 minutes of processing, you receive an email with your builder profile. So far, 551,126 sessions have been uploaded and analyzed through the platform.

Your Builder Profile

The profile breaks down how you work across five dimensions: steering, execution, engineering, product instinct, and planning. Paxel then assigns you an archetype based on patterns it sees in your sessions—Architect, Quality Guardian, Velocity Machine, or Night Owl. The system also extracts what they call "decision patterns"—your signature moves when directing the AI agent—and identifies a "growth edge," specific suggestions grounded in your actual session data rather than generic productivity advice.

Privacy Trade-offs

YC is upfront about what's transmitted: transcript excerpts (prompts, responses, tool call snippets) go to their proxy for summarization, and what actually reaches YC's servers is a small JSON payload containing scores, narratives, redacted decisions, and session metadata. Your working tree and .env files never leave your machine. You can read the full breakdown on their site if you want to verify their claims before trusting them with your data.

Why This Matters

Nobody really knows what "building well" means in the age of AI coding agents. Are you too hands-off, letting the agent drift off into boilerplate hell? Too micromanage-y, defeating the whole point of delegation? YC is essentially running a massive behavioral study to figure out what patterns correlate with good outcomes—and by uploading your sessions, you're contributing to that research while getting personalized insights in return. There's even a Paxel token you can attach to your Startup School 2026 application to boost your chances.

Key Takeaways

  • Free tool that analyzes your local AI coding transcripts from Claude Code, Codex CLI, and Cursor
  • Runs entirely locally via Docker—no code leaves your machine during processing
  • Delivers archetype classification (Architect, Quality Guardian, Velocity Machine, Night Owl) plus actionable growth recommendations
  • Over 551K sessions analyzed so far, with YC using the aggregate data to understand what "good AI-assisted development" looks like

The Bottom Line

Paxel is either a genuinely useful self-improvement tool or an elaborate experiment to crowdsource the definition of good AI coding practices—probably both. Either way, if you're shipping code with AI agents and have no idea whether your approach sucks, running this analysis once seems worth 20 minutes of your time.