AI is already writing your code. You just don't see how.
ARX captures the prompts, decisions, tool usage, and workflows behind every AI-assisted session —
so you can see what happened before it reaches production.
Git shows the outcome. ARX shows the process.
See it in action »
·
Install in 30 seconds
·
Report a Bug
Warning
ARX is currently in private beta. The parser has been tested against 1636 sessions, but your patterns will likely surface edge cases we haven't seen yet. Bugs at this stage are expected and useful please report them here.
AI is writing more of your code every day. But the moment the session ends, the process disappears.
You see the diff. You don't see why the model made the choices it did, which prompts sent it in the wrong direction, where it spun its wheels retrying the same failing tool call, or how many tokens you burned on context it already had.
Git records what changed. Nothing records how it got there until now.
ARX hooks into your AI tool's session context and captures everything: every prompt, every tool call, every decision, every token. When the session ends, it turns that into structured reports you can act on.
Works with: Claude Code · OpenAI Codex · Cursor · GitHub Copilot
curl -fsSL https://get-arx.ber.run/install | bashInstalls the arx binary to ~/.local/bin (or the first writable directory on your PATH). Available on Linux, macOS, and Windows.
The installer auto-detects your tools. It checks for Claude Code, Cursor, and Codex and automatically configures hook integrations for each one it finds. No manual config required.
Then use your AI tools as normal. When you're done:
arx timelineThat's it. ARX never blocks or modifies your sessions. If a hook fails for any reason, it fails silently and logs in the background.
To remove ARX and all its hooks cleanly:
arx uninstallNew to ARX? docs/quickstart.md walks through what just happened and what to run next.
Table of Contents
- Scorecard a session quality report: tool success rate, prompt clarity, token waste, and actionable improvement suggestions
- Timeline a full chronological audit log of every tool call, token cost, and agent decision
- Instruction compliance tracks whether agents follow
AGENTS.mdinstructions turn-by-turn, giving you a measurable signal where you'd otherwise have none
A high-level view of how a session went at a glance: execution (tool calls, failures, time split), prompt quality, cost efficiency, and concrete insights to improve the next session. Run it on any branch:
arx scorecardFull reference: docs/scorecard.md
A full chronological audit log of every tool call, permission request, and sub-agent — with token cost and context window usage per prompt block. Where the Scorecard tells you what, the Timeline tells you why and how.
arx timelineFull reference: docs/timeline.md
Instructions in AGENTS.md are not always followed, and there's no built-in way to know which sessions or turns skipped them. ARX tracks this with a lightweight ping mechanism the agent calls at the end of each turn, giving you a per-model compliance rate over time. Wired in automatically at install — nothing to add to your own AGENTS.md.
Full reference: docs/ping.md
ARX hooks into your AI tool's session context — not your IDE, not your network traffic — and records the structured data that your tools already produce. After the session, it uses your local AI tools to evaluate and summarize what happened. Nothing leaves your machine except your email and GitHub username, submitted once at install time so we know who's in the beta.
Full reference: docs/how-it-works.md
Your work stays on your machine. ARX never transmits your code, diffs, prompts, tokens, secrets, or API keys to us or anyone else — prompt analysis runs entirely through your local AI tools, which already hold that data.
Full reference: docs/privacy.md · docs/paranoid-setup.md for blocking all network access at the OS level.
Found a bug? Open an issue
Have a question or idea? Start a discussion
Security issue? Email dominis@ber.run do not open a public issue.