YC Wants to Know How You're Building With AI Right Now Y Combinator just dropped Paxel, an experimental tool designed to answer a question nobody's really asking yet: what does it actually look like when developers build with AI agents? The project emerged from the observation that while everyone's using coding assistants like Claude Code, Codex CLI, and Cursor, there's no collective understanding of how these workflows differ person to person. So YC built something to change that—and 552,010 sessions have already been uploaded for analysis.
How Paxel Works: One Command, Local Docker Getting started is straightforward but requires some setup. You need Docker running on your machine, then sign in with your email and execute a curl command from the repo you want analyzed. The script runs entirely locally in a Docker container, processing your AI session transcripts over 15 to 30 minutes depending on how many sessions exist across your projects. For broader analysis, you can run it from a parent directory containing multiple repos, or target individual projects with specific flags. A helpful prompt is provided for Claude, Codex, or Cursor that scans common transcript storage locations—~/.claude/projects/, ~/.codex/sessions/, and ~/Library/Application Support/Cursor/User/workspaceStorage/ on macOS—to find all your AI-coded repositories at once.
Your Builder Profile: Archetypes and Growth Edges The analysis produces a detailed builder profile scored across five dimensions: steering, execution, engineering, product instinct, and planning. Paxel identifies your archetype from four types—Architect, Quality Guardian, Velocity Machine, or Night Owl—based on patterns in how you direct AI agents during development sessions. The tool extracts your decision patterns and signature moves when working with coding assistants, pulling examples directly from your transcripts rather than generic observations. Most interestingly, it surfaces a "growth edge"—specific things to try next that are grounded in your actual session data instead of boilerplate productivity advice.
Privacy: Your Code Never Leaves Your Machine YC was clearly aware that asking developers to upload anything resembling their codebase would face immediate resistance. The architecture addresses this directly: your working tree and .env files never leave your machine. During analysis, transcript excerpts (prompts, agent responses, and tool call snippets) are sent via proxy to Claude or GPT for summarization and scoring. What's uploaded to YC is a small JSON payload containing only scores, narrative summaries, redacted decision patterns, and session metadata. Users can review the full breakdown of exactly what data travels where before committing anything.
Startup School Integration and Open Participation The tool isn't just about self-knowledge—there's strategic value for Y Combinator applicants. Linking your Paxel token to a Startup School 2026 application apparently boosts admission chances, giving builders direct incentive to try the platform. The project also accepts bug reports and feature requests through an integrated form, with an interesting twist: for features they like, YC will run prompts through a coding agent and merge promising results directly into the codebase.
Key Takeaways
- Paxel is free and runs locally via Docker—no code upload required, only transcript excerpts go to AI APIs
- 552,010 sessions analyzed so far, making this one of the largest datasets on real-world AI coding patterns
- Profiles aggregate across multiple machines when you log in with the same email
- Startup School applicants can attach their Paxel token directly to boost admission chances