Datasets:

pzarzycki
/

agents-claude-md-github-census-2026

pretty_name: AGENTS.md and CLAUDE.md in popular GitHub repositories (snapshot 2026-10-16)
license: other
license_name: per-repository
license_details: >-
  File texts keep the license of the repository they come from (column
  `license`). See the Licensing section.
language:
  - en
size_categories:
  - 10K<n<100K
tags:
  - agents-md
  - claude-md
  - coding-agents
  - agent-instructions
  - github
  - software-engineering
configs:
  - config_name: files
    data_files: files.parquet
    default: true
  - config_name: repos
    data_files: repos.parquet
  - config_name: population
    data_files: population.parquet

AGENTS.md and CLAUDE.md in popular GitHub repositories

A census of the agent context files /AGENTS.md and /CLAUDE.md in 22,854 popular, active open-source GitHub repositories. The snapshot was taken on 2026-10-16. This is the core data of the study github-coding-metastudy by Pawel Zarzycki: an arXiv paper and an interactive companion website.

Every repository in the population was checked; this is a census, not a sample. 6,006 repositories (26.3%) have at least one of the two files: 2,566 have AGENTS.md only, 1,052 have CLAUDE.md only, and 2,388 have both.

Paper

This dataset accompanies What Open-Source Projects Tell Their Coding Agents: A Census of AGENTS.md and CLAUDE.md.

Population

At the snapshot date, a repository is in the population if all of the following hold:

it is public on github.com;
its primary language (GitHub Linguist) is one of Python, JavaScript, TypeScript, Java, C#, C++, C, Go, Rust, PHP, Ruby, Kotlin or Swift;
it has at least 1,000 stars;
it was pushed within 90 days (pushed_at >= 2026-07-18, a push to any branch counts);
it is neither a fork nor archived.

Files

The study covers only /AGENTS.md and /CLAUDE.md, at the repository root, with exact case, on the default branch, in their current version (the commit is in head_oid). Other formats are out of scope, for example .cursor/rules, .github/copilot-instructions.md and nested AGENTS.md files.

Configs

All three configs join on full_name (owner/repo).

`files` (8,394 rows): one row per file

column	meaning
`full_name`, `path`	repository and file (`AGENTS.md` or `CLAUDE.md`)
`license`	the repository's license, an SPDX id from GitHub; `NOASSERTION` means GitHub couldn't identify it, null means no license file
`head_oid`, `head_date`	the default-branch commit the file was read from, and its date
`is_symlink`, `link_target`	whether the file is a symlink, and its target path
`byte_size`	size in bytes as reported by GitHub
`text_status`	`included`, `withheld_license`, `unresolved_symlink` or `binary`
`text`	the file content (UTF-8); null unless `text_status = included`

Symlinks are resolved, so a symlink row carries the target's text. Of the 874 symlink rows, most are a CLAUDE.md that points to the same repository's AGENTS.md. Filter on is_symlink to avoid counting that content twice.

Text status in this release:

text_status	count
included	8,387
unresolved_symlink	6
binary	1

Included texts: 79.0 MB.

`repos` (6,006 rows): one row per repository with at least one file

Repository metadata from GitHub at enrich time:

stars, forks, watchers, disk_usage_kb, code_bytes;
languages (bytes per language), topics, license, description, homepage_url, default_branch, created_at, pushed_at, is_fork, is_archived, is_template;
releases, open_issues, closed_issues, open_prs, merged_prs, commits_total.

Activity metrics:

commits_12m: default-branch commits in the 365 days before the snapshot;
recent_*: computed from the latest (up to) 100 of those commits. They are recent_sample (how many commits were sampled), recent_authors, recent_human_authors (bots excluded), recent_bot_commits and recent_ai_commits (commits with a co-author trailer or signature of Claude, Codex, Copilot, Cursor, Devin, Aider or Gemini).

How the two files relate:

relation: one label per repository. The rules are applied in order and the first match wins: symlink, thin @AGENTS.md import (300 characters or fewer), short pointer, import plus own content, identical copies, separate files;
relation_detail: for example, the symlink target.

relation	count
AGENTS.md only	2,549
CLAUDE.md only	1,040
both: CLAUDE.md is a thin @AGENTS.md import (<=300 chars)	771
both: CLAUDE.md is a symlink to AGENTS.md	629
both: separate files	262
both: separate files, CLAUDE.md mentions AGENTS.md	171
both: AGENTS.md is a symlink to CLAUDE.md	155
both: CLAUDE.md is a short pointer to AGENTS.md	140
both: CLAUDE.md imports @AGENTS.md and adds content	101
both: separate files, AGENTS.md mentions CLAUDE.md	64
both: identical copies	40
both: symlinks to the same other file	28
both: AGENTS.md only imports or points to CLAUDE.md	27
AGENTS.md only (symlink to another file)	17
CLAUDE.md only (symlink to another file)	12

`population` (22,854 rows): every repository that was checked

column	meaning
`full_name`, `repo_id`	repository
`primary_language`, `stars`, `created_at`, `pushed_at`	from GitHub search at discovery time
`found`	the repository still resolved when presence was checked
`has_agents_md`, `has_claude_md`	whether each root file exists
`agents_md_size`, `claude_md_size`	file size in bytes, if the file exists

Use population for adoption rates, for example by language or star band. Note that population.stars and population.pushed_at come from search time, while the repos columns come from enrich time a few hours later, so the values can differ slightly.

Collection

All data came through the GitHub GraphQL and REST APIs with one token, on 2026-10-16; no repository was cloned. The pipeline's code and its parameters are in the study repository (metastudy/).

Known limitations:

103 adopters have a default-branch HEAD older than the 90-day window. They passed the filter because of pushes to other branches.
Stars and pushed_at are a snapshot, so repositories near the 1,000-star cut-off move in and out over time.
A repository is classified only by its primary language.
7 files have no text: 6 symlinks can't be resolved (their targets are outside the repository or not in it), and GitHub flags 1 file as binary.

Licensing

The file texts are third-party content. Each text keeps the license of the repository it comes from, given in license. Check that license, and attribute the repository (full_name, head_oid), before you reuse a text.

This release includes every text that was collected. 1,358 included texts come from repositories with NOASSERTION or no license; their reuse terms are unclear.

Repository licenses among adopters:

license	count
MIT	2,334
Apache-2.0	1,525
NOASSERTION	816
GPL-3.0	389
AGPL-3.0	347
(none)	157
BSD-3-Clause	131
GPL-2.0	98
MPL-2.0	50
BSD-2-Clause	38
LGPL-3.0	31
CC0-1.0	12
ISC	11
LGPL-2.1	8
Unlicense	8
CC-BY-SA-4.0	7
EPL-2.0	7
CC-BY-4.0	6
Zlib	6
EUPL-1.2	5
MIT-0	4
OSL-3.0	3
0BSD	3
BSL-1.0	2
EPL-1.0	2
MS-PL	1
WTFPL	1
OFL-1.1	1
UPL-1.0	1
ECL-2.0	1
PostgreSQL	1

The metadata columns (counts, flags, labels) are facts about public repositories, compiled by the author.

Personal information

The texts are published files from public repositories. Some contain names, GitHub handles or e-mail addresses of maintainers: 302 of the included texts contain at least one e-mail address. They were not removed. Please don't use this dataset to contact or profile individuals.

Citation

If you use this dataset in your research, please cite:

@misc{zarzycki2026codingagents,
  title         = {What Open-Source Projects Tell Their Coding Agents:
                   A Census of {AGENTS.md} and {CLAUDE.md}},
  author        = {Pawel Zarzycki},
  year          = {2026},
  eprint        = {2610.31715},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SE},
  url           = {https://arxiv.org/abs/2610.31715}
}