Skip to content

AI & Co-authors ​

This lens answers: how much of the code did AI help write, where, and who started using it when? It also shows which people are credited together on commits.

It can only count AI that left a mark in git. A commit counts when an AI tool is named in a Co-authored-by: trailer, or when an AI agent made the commit itself. AI help that leaves no trace, such as editor autocomplete, is invisible here, so read every number as "at least this much."

Reading the numbers ​

What you seeWhat it means
"60% of commits had AI in them (900 of 1,500)"Of the commits by people and AI agents in the analyzed period, the share with AI in them. Bot commits and merge commits aren't counted on either side.
"the first in Jan 2026"The first month with an AI commit in the analyzed period, which is the last 12 months by default. It isn't when the team first used AI: earlier use falls outside the period. Use --since all to look further back.
"Claude 780, GitHub Copilot 120"Commits per tool. A commit naming two tools counts once for each, so these can add up to more than the total.
"Most AI commits: Jo Tan 340 of 400"The people with the most AI-assisted commits, out of each person's own commits. AI agents aren't people, so they get their own line.
AI-assistedA person made the commit, and a trailer names an AI tool.
Agent-authoredAn AI agent made the commit itself, such as the Claude or Copilot GitHub apps, Devin or Jules.
A package row's shareThe share of that package's commits with AI in them. Rows won't add up to the headline: a commit touching two packages counts in both rows, and some commits are left out of rows entirely (see How it's measured).
"services/api has the most AI"The highest share among packages with 20 or more commits, so a package with 5 commits, all with AI, won't take the title.
"0% of commits had AI in them (16 of 8,000)"Not a contradiction: the share is rounded. It's a very small amount of AI, and the count is the number to trust.
"34 pairs of people are credited together, 29 of them once"Pairs of people named on the same commit. When most were credited only once, the lens says so, because that's usually GitHub squash-merging a PR with several authors rather than pair programming. AI tools are never part of a pair.

The table's tiers are by a package's share with AI: No AI, Under 20%, 20–49% and 50% or more. They describe how much, and don't mean good or bad.

Three example repositories ​

These repositories are invented, but each result is what GitRelic reports for commits like these.

A Go backend with no AI showing ​

No commit had AI in it: no Co-authored-by trailer names an AI tool, and no AI agent committed (1,200 commits).

34 pairs of people are credited together, 29 of them once: mostly GitHub squash-merging a PR with several authors, not pairing.

What it means: nothing in git says AI was used. That doesn't prove nobody used it: the team may use tools that don't add a trailer, or their merge settings may remove trailers. If you know the team uses AI, this result says the history doesn't record it.

The pairs aren't evidence of pair programming either, as the second line says. The 5 pairs credited more than once are the ones worth a closer look.

A Python repo where one person started early ​

6% of commits had AI in them (120 of 2,000), the first in Sep 2026: Claude 112, Cursor 8.

Most AI commits: Priya Nair 102 of 160; Tom Okafor 12 of 240; Lena Voss 6 of 310.

services/ml has the most AI: 102 of 160 commits (64%).

What it means: 6% undersells what's happening and also oversells it. It's a 12-month average, and AI only appears in the last few months, so the recent share is higher. But most of it is one person: 102 of the 120 commits are Priya's, and the package with the most AI is the one Priya works in.

What to look at: Adoption lanes shows whether other people are picking it up. To see the current share rather than the yearly average, run again with --since "3 months ago".

A Java and TypeScript repo with AI throughout ​

60% of commits had AI in them (900 of 1,500), the first in Jan 2026: Claude 780, GitHub Copilot 120.

Most AI commits: Kim Berg 360 of 900; Jo Tan 340 of 400.

AI agents authored 200 commits themselves (GitHub Copilot 120, Claude 80), counted apart from AI-assisted ones.

services/api has the most AI: 420 of 480 commits (88%).

Changes in AI commits were fixed within 30 days more often than chance explains: 96 of 2,400 changes (4%) against 40 of 2,000 (2%) for the rest. AI commits are often the newest code, which changes more anyway: a reason to look, not a verdict.

What it means: AI is now a normal part of how this team works. The 900 is 700 AI-assisted commits plus the agents' 200. Jo uses it on 85% of their commits and Kim on 40%, so even here habits differ.

What to look at: the last line is the one to follow up, and it isn't saying AI code is worse. Newer code tends to change more whoever wrote it, and AI commits are often the newest. Check Fix rate to see which files those fixes landed in. When the gap is within chance, or there are too few changes to compare, the lens says that instead.

How it's measured ​

GitRelic reads the commits in the analyzed period and sorts each one by who made it, using the email address:

  • A bot's commit is left out of the share, the people and the package rows.
  • An AI agent's commit is agent-authored, whatever its trailers say.
  • A person's commit is AI-assisted if any Co-authored-by: trailer names an AI tool. Several AI trailers still make one AI-assisted commit.

The headline share is AI-assisted plus agent-authored commits, out of all commits by people and agents, rounded to a whole percent.

Trailers. Only Co-authored-by: counts, in any capitalization. Signed-off-by: and phrases like "with help from" don't. A bot named in a trailer is ignored, and so is crediting yourself, including under another email that resolves to you.

Package rows have stricter rules than the headline:

  • A commit counts once in each package it touched.
  • Commits touching more than 20 files are left out, because they say nothing about one package. They still count in the headline.
  • Only files that still exist in the analyzed commit count, so work on files since deleted or renamed doesn't appear.
  • A package is one of the repository's own declared workspaces, else packages/*, apps/*, libs/* or services/*, else the top-level folder.

The fix comparison uses Fix rate's changes: it compares how often changes in AI commits were fixed within 30 days with the rest, and warns only when the gap is beyond chance.

What counts as AI ​

ToolEmail address
Claude (Anthropic)noreply@anthropic.com, and the Claude GitHub app claude[bot]@users.noreply.github.com
GitHub Copilotcopilot[bot]@*.noreply.github.com, and its coding agent <id>+Copilot@users.noreply.github.com and copilot-swe-agent[bot]
Aideraider@aider.chat
Devin (Cognition)devin-ai-integration[bot]@*.noreply.github.com
Cursor*@cursor.sh, and Cursor's agent cursoragent@cursor.com
Amp (Sourcegraph)amp@ampcode.com
Jules (Google)google-labs-jules[bot]@users.noreply.github.com

A GitHub app's address may start with a numeric id (209825114+claude[bot]@…); each pattern accepts it. If your team uses an AI tool that isn't listed, please file an issue with its email address.

What counts as a bot ​

Addresses starting dependabot@, renovate@ or semantic-release@ (or those names with [bot], or semantic-release-bot@), github-actions[bot]@, and any other [bot]@*.noreply.github.com account that isn't on the AI list above.

In the dashboard ​

Find this lens under Team & Activity. Little or no AI is an answer too: the picture is the same, nearly all grey, and says so.

  • Where AI writes is the main picture. Each package is a row, most commits first, with its bar as long as its commits. The accent part is AI-assisted commits, and the lighter part is agent-authored. The share is at the right. Up to 18 rows are drawn, and the rest are in the table. Select a package to bring its row forward, or a person to redraw the rows from their own commits. Hover a bar for its counts.
  • Adoption lanes (the corner toggle) has one row per person, showing their commits over time with the part with AI in the accent. Everyone with an AI commit comes first, and AI agents get their own rows, followed by the busiest of the rest, up to 14 rows. The note at the right gives their share and first AI commit, such as "10% · since Jun 29."
  • The time strip under the picture shades each week's commits with AI in them. With a selection, it shades the selection's.
  • The inspector, with nothing selected, shows the findings quoted above. A person's card adds how many of their commits were AI-assisted and who they're credited with most, when they have either.
  • The table (T) lists every package: commits, how many had AI in them, the share, and the AI-assisted and agent-authored counts.

What it can't tell you ​

  • How much of the code AI wrote. A trailer says AI helped with a commit, not how much of it. One line and a whole feature count the same, and a person chose to leave the credit, or their tool did.
  • Whether people work together. Being credited together isn't the same as coordinating, and many teams who pair never add trailers. See Parallel Dev for people editing the same files in the same week, with or without credit.
  • About tools it doesn't recognize. The tool list is exact email addresses, so a new tool counts as a person until it's added.
  • What your merge settings removed. Squash merges can keep or drop trailers depending on how the repository is set up. A repository that drops them under-reports both AI and pairs.
  • Fix rate — how often changes need a fix of their own lines, behind the comparison above.
  • Parallel Dev — people editing the same files in the same week, with or without credit.
  • Contributors — who works on what, and when.
  • Ownership — how much of each package rests on one person. It counts commit authors, not co-authors, so a person working with AI looks like one person there.

Released under the MIT License.