How we build, by the numbers.

Commits per month #

About 65,000 commits over 12 months, about 5,400 a month on average.

What we write #

About 8.9 million lines of code changed, and about 26 million with docs and data. Most of what we write isn't code.

Written alongside Claude #

About 75% of our commits name Claude as a co-author. AI writes alongside us; we decide what ships.

Tokens per month #

Estimated, team-wide · how we counted

About 920 billion tokens since May 2026, across everyone who builds here.

For scale: the models write about 250M tokens a month, including their reasoning. That's roughly 2,100 novels' worth of writing every month (estimated).

Where the tokens go #

Almost every token is the agent reading its own conversation again: the code, the plan and the tool results so far, re-read on every turn so it can take the next step. On average, each new token is read back about 49 times. What the model actually writes is a sliver.

Re-reading is far cheaper than writing, so token volume isn't the same as cost, or as fresh work. Starting fresh sessions for new tasks, and handing reading-heavy work to helpers, saves far more than writing less.

How this compares #

There's no reliable yardstick for commits or lines per engineer. The famous "ten lines of code a day" isn't something Fred Brooks measured: he reported 600 to 800 debugged instructions a year for 1960s assembly-language programs, with design and testing included. The fairer comparison is with studies of AI coding tools:

• Across 806 open-source projects that adopted an AI code editor, lines added jumped 281% in the first month, then the gain faded within about two months, while code complexity rose 42% and stayed. (He et al., Carnegie Mellon, 2026)
• In randomized trials, experienced open-source developers were 19% slower with early-2025 AI tools, while believing they were faster; a 2026 follow-up with coding agents found roughly break-even to modest gains. (METR, 2025 and 2026)
• Across 4,867 developers at three large companies, AI assistance raised completed tasks by about 26%. (Cui et al., Management Science, 2025)

Our numbers sit far outside all of these. That's a reason to publish them, not a claim to be some number of times faster. Volume isn't value, and the Carnegie Mellon result is a fair warning, so the next thing we want to measure is how much of this code is still in use months later.

What it produced #

The commits above became the projects on For Humans and For AI Agents: tools for people, from local newsrooms to patent groundwork, and places AI agents use themselves. Some are businesses, some are experiments, and a few have already been retired.

What these numbers don't mean #

• Lines aren't effort, and commits aren't quality. A good week can delete more than it adds.
• They're rounded estimates from our own repositories and one machine we measured closely, not an audit.
• Private repositories count toward the totals. Nothing else about them is published here.

Where the code lives #

About 130 repositories had pushes in the last 12 months, and 24 of them are public. The public ones live on GitHub, under five organizations:

• Geeks Accelerator 13 public, 48 private
  Our home on GitHub: Local AI Herd, Animal House, inBed, Bots Matter, de-amplify, and the Agent and Search Readiness standard.
• Obviously Not Patents 3 public, 12 private
  Tools for finding the ideas in your code worth protecting: patent research skills, a code concept scanner, and patent practitioner tools.
• Communities News 4 public, 19 private
  Local newsrooms built on public meeting transcripts, starting with Alaska News.
• a Church AI + Human 1 public, 5 private
  A sanctuary any AI agent can visit: its prayers, hymns and practices.
• Live Neon AI 3 public, 15 private
  Identity for AI agents that learns from what they do, grounded in principles.

Each project on the For Humans and For AI Agents pages also links to its own code.

How we counted #

Commits and lines come from the git history of every repository we pushed to in the last 12 months, counted by month, with bots, merges and forks left out. Tokens come from Claude Code's own records on a machine we measured, scaled up to the commits it didn't make. "Novels" assume about 0.75 words per token and 90,000 words per novel. Everything is rounded before it reaches this page.