stillthere

How much of what your AI agent wrote is still in your repo?

npx stillthere

  WHO WROTE IT              COMMITS     WRITTEN    STILL THERE   SURVIVAL
  ────────────────────────────────────────────────────────────────────────
  ▸ Claude                    158    25,347      21,098       83%
  · You                        98     4,603       4,360       95%

  Your own code outlives agent code by 11 points.

Every usage tool on the market answers what did it cost. This one answers what was it worth — of the lines your agents actually wrote, how many are still in the file today.

Why nobody else ships this

Vendors won't. A tool that reports "38% of what Claude wrote for you is already gone" is not a tool Anthropic, OpenAI or Cursor will ever build, because the number is only interesting when it's unflattering. Every other measurement in this space — cost, tokens, rate limits — is safe to publish. This one isn't.

Meanwhile the phenomenon is well documented in aggregate: research across 153M lines found ~40% of AI-written code is rewritten or deleted within two weeks, and GitClear measured churn doubling since 2021. What did not exist was the personal version: not the industry average, but your repo.

How the number is made

Authorship comes from trailers the agents stamp themselves — Co-Authored-By: Claude, Co-authored-by: Copilot, Generated with Codex, and the equivalents for Cursor, Devin, Jules, Gemini and Aider. Nothing is inferred from writing style, because style attribution is guesswork and guesswork can't carry a claim like this.

Survival comes from git blame at HEAD. Lines written are read from git log --numstat; lines surviving are the ones blame still attributes to that commit. Survival is one divided by the other.

Run npx stillthere --json for the raw numbers.

What it refuses to do

It will not print an impossible number. If blame credits more surviving lines than history says were written, the tool says attribution didn't close and reports nothing. That failure is real: it happens with heavy file moves, and it happened on the first repo tested here — a 35,486-line demo.gif that git counts as binary but blame reads as text. Binaries and lockfiles are now excluded, and the guard stays in for the cases nobody predicted.

It will not guess. A commit that merely mentions an agent in prose ("drop the claude-specific workaround") is human-authored, and is counted that way.

It does not measure quality. A deleted line is not a bad line — code gets removed for a dozen good reasons. This measures durability, which is one honest signal among several, not a verdict on the tool that wrote it.

Options

npx stillthere [path]        survey a repo (default: current directory)
  --since "3 months ago"     limit to recent history
  --json                     machine-readable output
  --include-lockfiles        count lockfiles (off by default — they are noise)

Install

npm i -g stillthere

Requires Node 20+ and git. Everything runs locally: no account, no API key, no network call. The tool reads your repository and prints to your terminal.