How AI coding tools sign the commits they help write

Git has no field for “an AI helped write this”. What exists instead is a set of conventions that tool vendors adopted independently, mostly borrowed from the co-author trailer GitHub introduced for pair programming. If you know the conventions you can read AI assistance directly out of public history. This is that list, and more importantly, what each signal is worth.

The four signals

A co-author trailer. A line at the end of the message naming an address the vendor controls:

Co-Authored-By: Claude <[email protected]>

This is the strongest signal. It is deliberate, it survives cherry-picks and merges, and the address space is unambiguous: nothing but the tool writes [email protected]. Claude Code, Cursor’s agent, Codex, Aider, Continue, Cody, Devin and Blackbox all use some form of it.

A bot author address. Some agents commit as themselves rather than crediting a person, using a fixed address such as [email protected]. Just as precise as a trailer, but it means the commit has no human author to credit.

A bot author name. Tools that ship as GitHub Apps get a deterministic author name of the form name[bot] — devin-ai-integration[bot], google-labs-jules[bot], copilot-swe-agent[bot]. Matching is anchored, so a near miss fails rather than guessing.

A message-body line. A few tools write a fixed phrase into the body, like Generated with Claude Code or Made-with: Cursor. The weakest of the four: body text is free-form, so it is the signal most easily lost when a message is rewritten.

What a signal proves, and what it does not

A detected commit proves the tool touched it. It says nothing about how much of the commit the tool wrote. A one-character fix and a thousand-line refactor carry identical trailers.

The absence of a signal proves nothing at all. Three very common situations produce AI-assisted work with no trace whatsoever:

1. Inline completion. The dominant mode for Copilot, Cursor and Blackbox leaves ordinary commits authored by the developer. This is most AI-assisted coding, and it is invisible here.
2. Attribution turned off. Every trailer-writing tool can be configured not to write one, and plenty of teams do exactly that.
3. History rewriting. A squash merge collapses a branch into one commit and usually drops the trailers with it.

Add private repositories, and the honest summary is that these counts are a floor on disclosed AI assistance, not a measurement of AI assistance.

Why the ranking still means something

Because the thing being counted is consistent. Every commit in it was published in a public repository by someone who chose to leave the attribution in place. Comparing two developers who both leave trailers on is fair. Comparing a trailer-writing tool against a completion-first one is not, which is why the tool pages state each tool’s mode rather than putting them all on one bar chart.

The most common misreading of this data is treating a tool’s total as market share. It is disclosure share. Copilot is almost certainly the most used tool in the world and one of the smallest numbers here, because its main mode signs nothing.

Reading a commit yourself

Everything above is visible with git alone:

git log --format='%an <%ae>%n%b' -20 | grep -iE 'co-authored-by|generated (with|by)|made-with'

Run it on your own repository and you will see exactly what this site sees. If you would rather not appear, the same command shows you what to remove, and the methodology page covers removal from the site itself.