ONE COMMAND · NOTHING LEAVES YOUR MACHINE
Your agent wrote it.Is it still there?
Every AI usage tool tells you what the tokens cost. This one tells you what the code was worth — how much of what your agent wrote is still in the file today.
$ npx stillthere WHO WROTE ALIVE SURVIVAL ───────────────────────────────────── ▸ Claude 25,347 21,098 83% · You 4,603 4,360 95% Your code outlives the agent's by 11 points.
MEASURED TWICE
The same gap showed up in two unrelated repositories.
A25,347 agent lines — 83% survive, against 95% for human-written code.
B17,193 agent lines — 73% survive, against 84%. Different codebase, different age, the same 11-point distance.
WHY THE SLOT IS OPEN
No vendor will ever ship this measurement.
A tool reporting “38% of what Claude wrote for you is already gone” is only interesting when it is unflattering.
Cost, tokens and rate limits are safe for a vendor to publish. Durability is not.
The phenomenon is documented in aggregate — research across 153 million lines found roughly 40% of AI-written code rewritten or deleted within two weeks. What did not exist was your repository’s version of that number.
HOW THE NUMBER IS MADE
Trailers the agents stamp themselves, and git blame at HEAD.
01Authorship comes from Co-Authored-By: Claude, Co-authored-by: Copilot and the equivalents for Codex, Cursor, Devin, Jules, Gemini and Aider. Nothing is inferred from writing style.
02Lines written are read from the commit history; lines surviving are the ones blame still attributes to that commit.
03No account, no API key, no network call. It reads your repository and prints to your terminal.
WHAT IT REFUSES TO DO
It will not print an impossible number.
If blame credits more surviving lines than history says were written, it reports that attribution failed instead of inventing a rate. That case is real: the first repository tested here held a 35,486-line demo.gif that git treats as binary and blame reads as text, which produced a survival rate above 200% before binaries were excluded. It also never guesses — a commit that merely mentions an agent in prose is human-authored — and it does not measure quality, because a deleted line is not a bad line.