ai-written commits — tracked daily since 2025
THE AI CODE INDEX
Which AI coding tool is actually winning? Daily commit volume, market share, and doubling time across 7 tools on all public GitHub repos.
§ 1 — total daily commit volume, all tools
fig. 01
112M commits · last 6 months · ▲ 416% vs prior span
latest complete day — sep 26 · 1.2M
§ 2 — market share of ai-written commits
fig. 02
§ 3 — the race for #2
fig. 03
cursor
36.5k/day · ▼ 22%
codex
3.6k/day · ▲ 35%
aider
128/day · ▼ 16%
gemini cli
41/day · ▼ 57%
claude code owns the index, so it's out of this grid — one shared scale for everyone else.
§ 4 — momentum, ranked
fig. 04
| no. | tool | momentum 30d | commits/day | doubles every | share |
|---|---|---|---|---|---|
| 01 | claude codefastest | ▲ 128% | 1.2M | 24 days | 96.6% |
| 02 | codex | ▲ 35% | 3.6k | 68 days | 0.3% |
| 03 | aider | ▼ 16% | 128 | — | 0.0% |
| 04 | cursor | ▼ 22% | 36.5k | — | 3.0% |
| 05 | gemini cli | ▼ 57% | 41 | — | 0.0% |
01claude codefastest
▲ 128% · 1.2M/day · 24d to double · 96.6%
02codex
▲ 35% · 3.6k/day · 68d to double · 0.3%
03aider
▼ 16% · 128/day · — · 0.0%
04cursor
▼ 22% · 36.5k/day · — · 3.0%
05gemini cli
▼ 57% · 41/day · — · 0.0%
§ 5 — the pr arena
fig. 05
fig. 05 — autonomous agent prs, all time
github copilot and devin don’t sign their commits, so they score zero in the charts above. zero there isn’t absence — it’s invisibility. this is their arena.
§ 6 — the open model layer
fig. 06
models — release snapshots (big stars, low velocity)
- mistralai/mistral-vibeMistral Vibe is a free, open-source AI assistant that lets developers have a conversation with their codebase directly from their computer's command line, using plain English to explore, edit, and manage code projects. It's built by Mistral, a leading European AI company, and runs on their own AI models.22/100Active5.0k★ · 712⑂
- deepseek-ai/DeepSeek-V3DeepSeek-V3 is a powerful open-source AI language model — similar to the technology behind ChatGPT — developed by the Chinese AI lab DeepSeek. It can understand and generate text, answer questions, write code, and handle complex reasoning tasks, and is freely available for anyone to use or build products on top of.15/100Active105k★ · 16.7k⑂
- meta-llama/llamaThis is Meta's official release of Llama 2, a family of powerful AI language models that anyone — from individual developers to large companies — can download and run on their own systems. It provides the core model files and basic code needed to load and use these AI models, which range in size and capability, for tasks like answering questions, writing, and conversation.11/100Active59.6k★ · 9.8k⑂
- deepseek-ai/DeepSeek-R1DeepSeek-R1 is a family of open-source AI models that are exceptionally good at reasoning through complex problems in math, coding, and logic — trained using a technique called reinforcement learning, where the AI essentially teaches itself by trial and error rather than being hand-guided. DeepSeek claims these models perform on par with OpenAI's most advanced reasoning AI (o1), and they're freely available for anyone to download and use.3/100Active92.0k★ · 11.7k⑂
- zai-org/GLM-4.5—4.4k★ · 477⑂
runtime — where open-model energy lives
- MoonshotAI/kimi-codeKimi Code is an AI-powered assistant that lives in your computer's command line and acts like a tireless junior developer — it can read and modify your codebase, run programs, search the web, and make decisions on its own to complete tasks. It's built by Moonshot AI and connects to their Kimi AI models, though it can also work with other AI providers.54/100Hot7.7k★ · 1.3k⑂
- ollama/ollamaOllama lets anyone run powerful AI language models (like DeepSeek, Gemma, and Qwen) directly on their own computer, without needing cloud services or an internet connection. It works on Mac, Windows, and Linux, and comes with a simple interface so developers and non-technical users alike can chat with or build applications on top of these AI models locally.51/100Hot182k★ · 18.0k⑂
- ggml-org/llama.cppllama.cpp is a free, open-source tool that lets you run powerful AI language models directly on your own computer or server, without needing expensive cloud services or specialized hardware. It works across a wide range of devices — from MacBooks to gaming PCs to cloud servers — and supports aggressive compression techniques (called quantization) that shrink AI models so they run faster and use less memory.40/100Hot130k★ · 23.9k⑂
- unslothai/unslothUnsloth is a desktop app that lets anyone run and train AI models — including popular ones like DeepSeek, Gemma, and Llama — entirely on their own computer, without needing cloud services or technical expertise. It supports both text-based AI and image-generation models, all through a visual interface rather than requiring code.35/100Active77.0k★ · 7.1k⑂
the tools above run on these. open models don’t sign their commits, so this layer is repo traction — not usage share. anything else would be made up.
§ 7 — ecosystem adoption
fig. 07
config files in repos
sdk / package adoption
note: agents.md (an open format) beats every vendor-specific config file.
§ 8 — community pulse
fig. 08
fig. 08 — community pulse, 7 days
| tool | mentions | hn | github | |
|---|---|---|---|---|
| claude code | 591 | 72 | 0 | 519 |
| github copilot | 261 | 7 | 0 | 254 |
| codex | 178 | 0 | 0 | 178 |
| cursor | 36 | 0 | 0 | 36 |
| gemini cli | 31 | 0 | 0 | 31 |
| aider | 6 | 0 | 0 | 6 |
| windsurf | 4 | 0 | 0 | 4 |
§ 9 — intelligence brief
fig. 09
$ tail -f /var/log/gitfind/index.log
09:41:02 index sync completed — 1.2M commits counted across 5 of 7 tools
09:41:03 claude code leads the index at 96.6% of ai-written commits
09:41:04 claude code added 651k/day this month — biggest absolute gain
09:41:05 claude code is the fastest doubler — every 24 days at current pace
09:41:06 devin leads the pr arena with 2.1k agent prs — without signing a single commit
09:41:07 gemini cli down 57% this month — watching for a floor
09:41:08 █
form 27-b — subscription
THE TUESDAY BRIEFING
The repos that moved this week, why they matter, and what to watch next. One email. No noise.