About

How we classify commits and measure the impact of AI coding tools on GitHub — including public and private repositories.

How Commits Are Classified

Every commit is run through a priority-based cascade. The first matching rule wins, and no further checks are performed. This ordering ensures the most authoritative signals take precedence.

GitHub API type field — If GitHub marks the commit author as a Bot account, we classify it immediately. This catches Copilot Agent, Devin, Sweep, Gemini Code Assist, and others.
Author email patterns — Emails containing dependabot, renovate, or ending in [bot]@users.noreply.github.com are classified as bots. AI agent email domains (e.g. cursoragent@cursor.com) are also matched here.
Author login and name — The commit author's GitHub username and git author name are checked against a database of 40+ known bot and AI tool patterns.
Committer-based detection — Some bots commit through GitHub's web UI. When the committer is a bot account and the commit message first line matches a known bot pattern, it's classified accordingly.
Co-Authored-By trailers — AI tools like Claude Code, Cursor, and Copilot add Co-Authored-By trailers to commit messages. We parse these to identify AI-assisted commits and, when possible, the specific tool used.
Commit message markers — Messages containing phrases like "Generated with Claude Code", "Generated with Cursor", or the aider: prefix are classified as AI-assisted.
Author name suffixes — Some tools like Aider append (aider) to the git author name.

If none of these rules match, the commit is classified as human.

PR reclassification: After initial classification, we perform a second pass on squash-merge and merge commits. If a commit references a pull request, we check whether the PR was opened by a bot account, or if the PR body, branch name, or labels contain AI tool markers. This catches cases where an AI tool creates a PR but the merge commit shows the human who merged it as the author.

What Counts as "AI"

We distinguish between AI coding tools and automation bots. Only AI tools are included in the AI percentage metrics.

Based on these percentages, we assign each developer a Rank – from "Organic Architect" to "Digital Overseer". You can find the full breakdown of ranks on our Developer Ranks page.

To ensure your tools are correctly detected, see our AI Attribution guide.

AI Tools (included in AI %)

GitHub Copilot — Including Copilot Agent and SWE Agent
Claude Code — Anthropic's coding assistant
Cursor — Cursor editor and Cursor Agent
AI-assisted — Devin, OpenAI Codex, Aider, Windsurf, Gemini Code Assist, Amazon Q Developer, and other AI coding tools

Automation Bots (shown separately from AI %)

Dependabot — Dependency version updates
Renovate — Dependency version updates
GitHub Actions — CI/CD automation
Other bots — Vercel, Netlify, Codecov, semantic-release, imgbot, and similar automation

Automation bots like Dependabot perform routine maintenance — they don't represent the kind of AI-assisted coding that this project aims to measure. They are tracked and shown separately (as "Bot Commits" / "Automation") so you can see the full picture without inflating the AI percentage.

How Lines of Code Are Measured

After fetching commits from the GitHub REST API, we enrich them with line-of-code data using the GitHub GraphQL API.

Additions only — We use the additions field from GitHub's GraphQL API. Deletions are tracked but not used in percentage calculations.
Best-effort enrichment — If the GraphQL query fails (due to rate limits, missing tokens, or API errors), the pipeline continues without code-volume data. The UI gracefully falls back to commit-count mode. You may see a warning when code-volume data is incomplete.
3-way split — Code volume is split into human, AI tools, and automation bots — the same three categories used for commit counts. All three are visible on the heatmap and stat cards.

Private Repository Support

Users who sign in with GitHub can optionally link their private repository data. Here's how it works:

Aggregate counts only — We fetch commit metadata from your private repos, classify each commit in memory, and store only the daily and weekly totals (human / AI / bot counts). No repository names, commit messages, code, or file paths are ever stored.
Same classification pipeline — Private commits go through the exact same priority-based cascade described above. The results are merged into your profile's activity timeline and stats.
Code volume not available — The GitHub API does not return line-of-code data from the list-commits endpoint used for private repos. The heatmap's code volume view reflects public repos only.
Privacy controls — You can toggle whether visitors see your private data on your profile, or keep it visible only to yourself. You can also unlink all private data at any time, which permanently deletes the stored aggregates.

Leaderboard rankings are always based on public data only, ensuring a fair comparison across all users regardless of whether they've linked private repos.

Data Scope & Limitations

There are important differences between our data and what GitHub shows on a user's contribution graph.

Factor	AI vs Human (Public)	AI vs Human (Private)	GitHub
What's counted	Commits to default branch only	All commits to default branch	All contributions (commits, PRs, issues, reviews)
Repositories	Top 20 non-fork repos by stars	All private repos (up to 200)	All repositories the user has contributed to
Time range	Last 2 years	Last 2 years	Last 1 year (contribution graph)
Timezone	UTC (all dates bucketed at UTC midnight)	UTC	User's profile timezone
Timestamp used	`authoredAt` (when the code was written)	`authoredAt`	`committedAt` (when the commit was recorded)
Code volume	Enriched via GraphQL API	Not available	N/A

Because of these differences, the activity you see on AI vs Human will not match GitHub's contribution graph exactly. Our focus is on analyzing commit authorship — who (or what) wrote the code — rather than replicating GitHub's contribution tracking.

Open Source & Contributions

The classification logic is open source. If you find a misclassification or know of an AI tool we're not detecting, contributions are welcome.

View on GitHub