Engineering telemetry
Your AI workflow, measured.
One command reads the session logs your coding agents already keep and scores what they show: whether the work they shipped was checked, finished, and real.
Metadata only. Your code, prompts, commands and file paths never leave your machine.
npx @beon-tech/ai-scoreTakes about a minute. Your result is private until you share it.
The reading
One score, six dimensions
Leverage, craft, verification, delivery, consistency and customization — each read from what your sessions actually show, not from what anyone claims. Where it applies they follow Anthropic's AI Fluency framework: leverage is delegation, verification is discernment, and craft is diligence and description — carrying a failing check through to a fix, and prompts that say what you mean. The meters below show the example run above.
Thin evidence discounts a dimension instead of guessing — meters show rates × confidence, and the reasons are printed under your breakdown.
Procedure
Two commands, in order
Nothing to install and nothing to configure. The scan reads your recent real usage, so run it on the machine you actually work on — a fresh setup has no history to score.
- STEP 1 / 2
Sign in from your terminal
Opens a page with a code to approve, so the run is attributed to a verified account rather than to whatever email your terminal happens to be configured with.
npx @beon-tech/ai-score login - STEP 2 / 2
Run the scan
Reads the local logs your coding agents already write, scores them, and prints your result with a private link to the full breakdown — visible only to you.
npx @beon-tech/ai-score
Privacy
We collect metadata. Never your content.
The CLI reads your coding-agent logs on your machine, counts what happened, and sends only those counts. Nothing you wrote, nothing the agent wrote, and nothing about your code ever leaves your computer.
Never sent
- Your prompts, or anything else you typed
- The agent's replies
- Your code, diffs, or commit contents
- Commands the agent ran, their arguments, or their output
- File names, folder paths, or repository names
- Anything from a machine you didn't run the command on
Sent — counts and states only
- How many prompts you sent, how many words they had, and how many tool calls they triggered
- Which tools ran, and how often they errored or were denied
- Whether edits were checked, fixed after failing, and delivered — as yes/no states
- When sessions started and how long the longest autonomous turn ran
- Token totals per model
- Pull requests opened and lines added or removed
- Which coding agents were found and where they keep their logs (e.g. ~/.claude)