// the agent test, run on a transcript

Your leash

How many tool calls your agent made between two things you said. Scored on your machine, from your own log.

Run it on yours

Two ways in. Both score with the same file, and neither sends anything anywhere.

terminal
npx isitanagent

Reads your newest Claude Code session, prints the verdict and the receipts, and hands you a link. Zero dependencies. It opens no sockets.

no terminal
Drop a .jsonl session log or click to pick one · it stays in this tab

Usually in ~/.claude/projects/. Parsed line by line in your browser — 100MB logs are fine.

There is no upload endpoint in this site to upload to. The link the tool prints carries numbers in the URL fragment, which browsers never transmit to any server. Turn off your wifi and score a file — it still works.

Don't trust it. Check it.
npx isitanagent --dry-run

Prints the exact blast radius, then exits without reading a byte:

  • READS ~/.claude/projects/*/*.jsonl — and names the single file it would score
  • WRITES nothing. There is no write path in the tool.
  • NETWORK nothing. No fetch, no telemetry, zero dependencies.

Three ways to falsify that in under a minute: npm view isitanagent dependencies returns {}; grepping the installed package for network calls returns nothing at all; or unplug your network and run it — the output is identical.

01

The scale

02

What the corpus says

Recovery is reported as a count, never as a score. It sits near the ceiling in almost every session, so a percentage would hand everyone the same grade and measure nothing. The number that actually varies is the leash.