The most copied repository in the coding-agent ecosystem has 674 stars

The most copied repository I found has 674 stars. The repository with 281,485 stars came second.

I crawled 146,338 GitHub repositories and catalogued 760,652 extensions for coding agents: skills, subagents, slash commands, hooks, MCP servers, Cursor rules, instructions files, settings and plugins. Then I measured every one of them the same way. The ranking that came out of it disagrees with the one GitHub shows you.

How I measured

I took three numbers for every file, with the same method applied throughout.

Token cost first. I ran each body through a real tokenizer and applied each agent’s documented loading rules, because those rules differ. A skill announces itself with a short description and loads its body only when something invokes it. An instructions file loads whole, at the start of every session, whether you use it or not. The figures below are what a file adds to a session in which you have not yet typed anything.

Second, a static scan with a fixed rule set, which records the exact line that produced each finding. Grades run A to F.

Third, lineage. I compared every body against every other body by similarity and sorted the result into original, copy and undetermined. Across the corpus that came out at 54.9% original, 11.1% provable copy or fork, and 34.0% undetermined. I come back to that third bucket at the end.

Stars measure attention, copies measure use

These are the repositories whose original files were copied into other people’s repositories most often.

Cyanheads wrote 34 original files. They were copied 3,726 times. That is roughly five and a half copies for every star the project has ever received. For obra/superpowers the same ratio is about one copy for every eighty stars, a difference of more than four hundred times between the two.

A star costs one click and means “I have seen this”. A copy means somebody took the file, put it in a repository they own, and let it run against their code. Those are different events. A well-known project collects a lot of people who are looking rather than installing, so its star count runs ahead of its use. A project that solves one narrow problem collects copies from the people who have that problem, and not many stars.

At the level of individual files the pattern repeats.

One hook and one instructions file account for most of the copying at the top. Neither is a framework, and both are single files that do one thing. That is the shape that tends to get pasted into a new project, often by someone who never opens the repository it came from.

What each of these things costs you per session

Here is the token cost, as a median, a 90th percentile and a maximum.

MCP servers, plugins, hooks and settings load nothing of their own file into the session.

An instructions file at 957 tokens and a skill at 53 are not the same kind of object, and the eighteen-fold gap in the median understates the difference. A skill’s body is conditional: it costs a line of description until something calls it, then the body once. An instructions file is unconditional. It sits in the context before the first word of your question, in every session, for as long as it stays in the repository.

That is what makes stacking them expensive in a way that is easy to miss. Twelve instructions files at the median is about 11,500 tokens spent before work begins. Twelve skills is about 640. Nobody picks those twelve in one sitting; they arrive one at a time, each reasonable on its own, and the bill is never itemised. The largest single instructions file I measured is 55,112 tokens, and it loads in full every time.

The second most copied file in the ecosystem is a CLAUDE.md. Instructions files are among the most copied things here and also the most expensive to keep.

What the scan found

The grades came out at A 95.5%, B 2.3%, C 1.7%, D 0.3%, E 0.1% and F 0.1%. In absolute terms, 495 files landed at F.

The other number worth stating is about licences rather than code. 45,449 of the 146,338 repositories, 31%, carry no LICENSE file at all, and under default copyright that makes them all rights reserved. A further 7,662 carry a LICENSE that neither GitHub’s classifier nor I could identify — custom text, dual licences, a vendor’s own terms — which is not the same thing as having none, and I am not going to guess at what those permit. Either way the reader of a copied file has no terms they can act on. Across the corpus, then, a third of the repositories this copying draws on offer nothing that plainly permits it, almost certainly without anyone involved intending anything by that.

What I would not claim

The copy figures are a floor, not a total. 34.0% of the corpus is undetermined lineage, which means my similarity comparison could not confidently place it either way. If those files resolved, the ranking above would move. I expect the broad proportions would hold, but I cannot show that from what I have.

The scan reports what it found and the line it found it on. It does not certify anything. An A grade means the rules matched nothing, which is a narrower statement than “this is safe”.

A copy is not an accusation. Forking is how this ecosystem propagates, most of these files exist to be reused, and vendoring somebody’s hook into your repository is the normal and intended way to use it. The count measures spread. Reading it as plagiarism would be a mistake, which is also why the licence figure matters more than the copy figure does.

The token numbers follow each agent’s documented loading rules as they stood when I measured, and those rules are young enough to change. Every figure here was taken on 4 September 2026, from a catalogue that re-crawls itself nightly, so the counts have already moved.

None of this was hidden. Every file was public, and the numbers only needed somebody to count the same way twice. But the ecosystem grew faster than its own record-keeping: authors do not know how far their files have travelled, and the people running those files do not know what they cost. Stars were the only signal on offer, so stars became the signal, and stars measure something next to the thing most people wanted to know. The data came out of a catalogue I keep at agentmods.dev, where the same figures sit against each individual file.

Every figure here comes from the catalogue at agentmods.dev, which re-crawls itself nightly. More writing · how the data is licensed.