Paxel

Paxel is a tool from Y Combinator that reads your local AI coding session transcripts and turns them into a personalized builder profile. It works with Claude Code, Codex CLI, and Cursor: you trigger it with a single terminal command, and the analysis runs inside a Docker container on your own machine. YC frames it as an experiment, since there isn't yet a shared picture of what it means to build well with coding agents.

What It Is

Paxel is a developer self-analytics tool that processes the session transcripts generated by Claude Code, Codex CLI, and Cursor. Running a single command in your terminal extracts patterns from your real AI-assisted sessions and returns a scored profile across five dimensions: steering, execution, engineering, product instinct, and planning. The stated goal is to help builders understand how they work with AI at a point when most people are doing it alone with no shared frame of reference.

How the Analysis Works

After you sign in with your email, you run the provided command from either a parent directory (to cover all repos) or a single project folder. The script locates the AI session transcripts that Claude Code, Codex CLI, and Cursor store locally. During the run, transcript excerpts — prompts, agent responses, and snippets of tool calls — are sent out for summarization and scoring. The privacy policy states this routing goes to Anthropic's Claude and OpenAI's GPT via Microsoft Foundry, with Google's Gemini used to create embeddings for semantic search. At the end, a JSON payload of scores, narratives, redacted decisions, and session metadata is uploaded to YC.

What Stays Local and What Leaves

This is the part worth reading closely for a tool that ingests your coding history. Paxel says your raw transcripts, full prompts, full agent responses, full tool outputs, working tree, and .env files never leave your machine. What does leave is a derived layer: prompt excerpts, truncated command text, per-session narratives, steering traces, behavioral scores, and decision records. The privacy policy notes this can include local file paths such as your home-directory structure. It also states that uploaded data may carry identifiers belonging to other people — names and email addresses of repository contributors drawn from commit metadata — and that you are responsible for having permission to share them. On the lighter side, the policy says Paxel does not sell or share data, runs no advertising or cross-site tracking, and sets a single session cookie. Use is restricted to adults.

What You Get Back

Each upload produces a report, and your profile aggregates patterns across reports over time. The profile includes:

• Archetypes: Named patterns such as Architect, Quality Guardian, Velocity Machine, or Night Owl, based on how you direct the AI
• Decision patterns: Signature moves pulled from real exchanges in your transcripts
• Scoring across five dimensions: Steering, execution, engineering, product instinct, and planning
• Growth edge: Specific things to try next, grounded in your own sessions rather than generic advice
• Behavioral stats: Prompt length, agent parallelism, commit timing, plan-mode frequency, and more

Who It's For

Paxel is aimed at software builders who regularly use AI coding agents and want a data-driven view of their own workflow rather than a sense based on memory. YC positions it as an open experiment and invites users to suggest features, which it says it will build with a coding agent and merge.

Setup Path

The tool requires Docker Desktop or a compatible runtime such as colima or OrbStack to be installed and running. Analysis takes roughly 15 to 30 minutes depending on session volume, and results arrive by email. You can aggregate several machines under one account by running the same command on each and signing in with the same email, and you can target specific repos from a parent directory or with the --project flag.