Claudeonomics

Weekly commit counts sourced from the GitHub Search API, bucketed by committer-date:YYYY-MM-DD..YYYY-MM-DD. Each agent is queried with the trailer its tooling actually emits. Historical (closed) weeks are frozen on first successful read; the current in-progress week is re-queried on each scheduled refresh (daily).

These are approximate counts. GitHub's Search API returns estimates for large result sets and marks them with incomplete_results: true. Claude Code weekly volumes are large enough to trigger that flag, but observed drift between consecutive reads is under 0.1%, so we display the value as authoritative and note it here.

"Share" is share of the tracked agents, not all of GitHub. The 99.5% on the KPI strip means Claude Code accounts for 99.5% of the weekly commits across the six agents on this dashboard. It is not Claude Code's share of all public GitHub commits — that denominator (all human + bot commits) isn't obtainable from the Search API; it would need GH Archive or BigQuery.

Two agent categories. We split the challenger charts by how the agent runs, because comparing a terminal tool to a cloud agent is apples to oranges.

• CLI agents run in your terminal. You invoke them, they edit files locally, you commit. Claude Code, Copilot CLI, Codex CLI, Gemini CLI.
• Cloud / async agents run on a service. You assign them a task (issue, ticket, PR review), they open a PR when done. Copilot Coding Agent, Codex (ChatGPT).

Counted (CLI agents):

• Claude Code — Co-Authored-By: Claude <noreply@anthropic.com>
• Copilot CLI — Co-authored-by: copilot-cli
• Codex CLI — Co-authored-by: openai-codex
• Gemini CLI — Co-authored-by: gemini-cli

Counted (cloud / async agents):

• Copilot Coding Agent — Co-authored-by: copilot-swe-agent (GitHub renamed the product but the bot account still emits the old trailer)
• Codex (ChatGPT) — Co-authored-by: chatgpt-codex-connector

Each line starts on the date the product first became publicly available — research preview, public preview, or open-source release, whichever came first. Not GA. We chose this over GA-only because preview-phase adoption is part of each product's real story; clipping to GA would erase the most interesting growth curves.

Caveat: our backfill starts May 5, 2025, so the chart can't go further back than that. Claude Code's research preview launched Feb 24, 2025 — its line starts at May 5 instead because that's where our data starts. The pre-May data exists on GitHub but isn't in our dataset. Verified launch dates below regardless:

• Claude Code — research preview Feb 24 2025; GA May 22 2025
• Codex CLI — open-source release Apr 16 2025; weekly counts were single-digit noise (with one 153-commit fork-pollution outlier on May 5) until Jul 28 2025 when sustained adoption began at 157+/week, so the line starts there rather than at product launch
• Codex (ChatGPT) — research preview May 16 2025; the chatgpt-codex-connector trailer string only appeared in commits starting Sep 1 2025, so the line starts there rather than at product launch
• Copilot Coding Agent — public preview May 19 2025; GA Sep 25 2025
• Gemini CLI — open-source launch Jun 25 2025
• Copilot CLI — public preview Sep 25 2025; GA Feb 25 2026

Note on CLI challenger volatility. The smaller CLI lines (Copilot CLI, Codex CLI, Gemini CLI) can spike week-to-week when a single large repo runs a scripted batch commit with the agent's trailer. Day-to-day movement on those lines reflects repo-level noise more than product adoption — read the multi-week trend, not individual weeks.

Not counted: Devin (clean author:devin-ai-integration query returns ~92 commits because most Devin activity is on private repos; the looser bot-trailer query is polluted by spam repos forging trailers). Windsurf / Cascade (max ~400 commits all-time — no trailer adoption). Aider, Cline / Kilo Code, CodeRabbit, Cursor, Amazon Q Developer, Jules — no stable public trailer that survives false-positive checks. Dependency bots like Dependabot and Renovate are excluded as they perform dependency updates, not code authoring.

This counts only public repositories and may undercount (private repos, non-standard trailers) or overcount (false-positive matches). Reproduce it yourself — single-week lookup for Claude Code:
Data sampled weekly. Inspired by AI Focus and SemiAnalysis.