The AI coding tool index

Which AI coding tools are actually used?

Every leading AI coding tool leaves a small mark on the commits it helps write — a “Co-authored-by” line in the git history. So we counted those marks across millions of real public commits. The result is lopsided: one tool has run away with it, and it is still pulling ahead.

Figures are estimates from GitHub commit search, as of 13 July 2026. How we measure ↓

Runaway leader

Claude Code

Around 159× the commit footprint of the next tool — two orders of magnitude ahead.

And still accelerating

≈ 40× in a year

Claude-authored commits went from ≈ 100K to ≈ 4.1M a week, year on year.

One tool has run away with it

Counting the same way over the same month (June 2026), Claude Code’s footprint isn’t a little bigger than the rest — it’s in a different league.

Claude Code, June 2026

≈ 6.8M

commits carrying its trailer (est.)

Next best (Devin)

≈ 43K

159× smaller

The race for second

Set Claude Code aside and the rest of the field is both far behind and closely bunched. Here they are to scale against each other (June 2026).

Devin≈ 43Kcommits
Copilot Agent≈ 17Kcommits
Cursor Agent≈ 15Kcommits
aider≈ 7.5Kcommits
OpenAI Codex≈ 5.1Kcommits
Gemini CLI≈ 1.8Kcommits

Bars are to scale within this group only — Claude Code would be roughly 159× the longest bar here. All figures are estimates.

The growth is the real story

The same search, one year apart, on the leader. The gap is so large that even a noisy estimate can’t hide it.

week of 7–13 Jul 2025

≈ 100K

commits / week

week of 6–12 Jul 2026

≈ 4.1M

commits / week

≈ 40×

in twelve months

The full field

Every tool stamps its own line into the commits it writes. We show the exact string we count and the June 2026 estimate, so the ranking is like-for-like.

#ToolMakerWhat we countJune 2026 (est.)
1Claude CodeAnthropicCo-authored-by: Claude≈ 6.8M
2DevinCognitionCo-authored-by: Devin≈ 43K
3Copilot AgentGitHubCo-authored-by: copilot-swe-agent≈ 17K
4Cursor AgentAnysphereCo-authored-by: Cursor Agent≈ 15K
5aideropen sourceaider (aider.chat)≈ 7.5K
6OpenAI CodexOpenAICo-authored-by: openai-codex≈ 5.1K
7Gemini CLIGoogleCo-authored-by: gemini-cli≈ 1.8K

How we measure — and what this can’t tell you

What we count

Agentic AI coding tools add a line like Co-authored-by: Claude to the commits they help write. We count how many public commits carry each tool’s line, using GitHub’s commit search. It is the same idea behind our language index: count what’s actually in the repositories, not opinions about it.

The numbers are estimates, and we won’t pretend otherwise

GitHub’s search returns an approximate count for large queries. Run the identical search twice and it moves a lot — we saw the same June total come back as 10.9 million and then 15.9 million seconds apart (a 45% swing). So we do two things: rank every tool over the same single month for a fair frame, and publish only the order and the direction of growth — both robust here because the tools sit orders of magnitude apart. An exact, reproducible version is on our roadmap and needs its own data pipeline.

This measures tools that sign their work

We can only see tools that leave a trailer — Claude Code, Devin, Cursor Agent, Copilot’s agent, Codex, Gemini CLI, aider. It is blind to plain GitHub Copilot autocomplete and to code pasted from a chat window, which leave no mark. Read this as “which agentic tools show up in commit history”, not “which AI is most popular overall”.

It ranks tools, not the models underneath

Cursor or a CLI can run different models under the hood, and the trailer names the tool, not the model. A model-by-model view is a harder, separate question.