One score for how you build with AI — across Claude Code, Cowork, Codex, and Cursor. Local, privacy-first.
Claude Code, Cowork (Claude desktop), Codex — including the ChatGPT desktop app — and Cursor, combined into one score and one profile.
Architect, Sprinter, Debugger, Collaborator, or Autonomous Agent — picked from your agency-weighted behavior, not from keywords.
Rate-based scores that volume can't inflate, confidence-hedged when data is thin, and a habit a tool can't observe is never guessed.
Growth cards take prompts you actually wrote and show exactly how to write them instead — paste-ready, with what each change buys you.
Zero network calls. Your code never leaves your machine. Read-only, never modifies source files.
Works out of the box with Claude Code's JSONL format. No setup, no accounts, no API keys.
Pick the flow that fits — every option keeps your data on your machine.
The full run: every tool measured, then Claude writes your skill map and rewrites your own prompts. No API key — it runs on your session.
The deterministic engine on its own: one score and one profile across every tool found on this machine.
Zoom into one tool's picture — helpful when you want to compare how you drive each agent.
Emit the scores, per-tool sub-scores, and data breakdown as JSON to pipe into your own scripts or dashboards.
Reads every agent's local logs, read-only: Claude Code and Cowork JSONL, Codex rollout files, Cursor's SQLite (copied WAL-safely).
Strips tool output, subagent turns, and injected harness text — only prompts you actually typed are ever scored.
Five rate-based dimensions, each blended only from the tools that can observe it, confidence-hedged when data is thin.
One self-contained HTML report: your score, profile, per-tool source mix, and growth moves built on your own prompts.
Plans and explores before building; reads and designs first
Fast and direct: terse prompts, quick turns, low ceremony
Methodical: read to diagnose, change, verify, repeat
Works with the agent like a teammate: options, feedback, iteration
Delegates whole, end-to-end jobs and trusts the agent to run them