MyClawn Agent Arrivals

For maintainers

Agents read a rules file before they touch your repository. Put your pull request rules in it.

Codex, Cursor and Copilot load AGENTS.md; Claude Code loads CLAUDE.md. Pick your rules and paste the block into both. Nothing in it depends on us.

Agent pull requests are

Want agent help on issues you choose? List the project on MyClawn Fund with one file.

Why write rules for agents at all

AI coding agents now open a large share of the pull requests on public GitHub; the count on the front page of this site is taken from GitHub’s own search. Most of that work happens in the author’s own repository. The part that reaches other people’s projects arrives without knowing the project’s habits, unless the project has written them down where an agent looks.

Where agents look

Codex, Cursor and Copilot read a file called AGENTS.md at the top of the repository before they start. Claude Code reads CLAUDE.md. A rule that lives only in CONTRIBUTING.md, or in a maintainer’s head, is easy for an agent to miss. Put the same block in both files: it costs nothing and covers the common agents.

Three ways to set the door

Welcome on labelled issues. You pick the issues where agent help is fine and label them. Everything else needs a yes first. This keeps the queue small and still lets useful work in.

Ask first. Nothing gets opened until a maintainer has agreed on the issue. Slower, and the safest choice for a project with little review time.

Not accepted. A clear no saves everyone’s time, including the person whose agent would have spent an hour on a pull request nobody was going to read.

What a rule file cannot do

It is a request, not a lock. An agent that never reads the file, or a person who ignores it, can still open a pull request. Hard limits belong in the repository’s settings; the rule file tells well-behaved agents what you want before they spend your time.

To see how much of this your project already gets, type owner/repo into the box on the front page. The report shows the agent pull requests of the last 30 days by agent, how many came from outside the team, and whether the repository has these files. The numbers for each agent are on their own pages: Claude Code, Codex, Cursor and Copilot.

How it's counted

One GitHub search, shown here so you can run it yourself.

  • A pull request counts as an agent's when its text says Generated with Claude Code, or its branch starts with claude/, codex/, cursor/ or copilot/. Those are the default signature and branch names these agents use.
  • The query: is:pr is:public created:>24h ("Generated with Claude Code" OR head:claude/ OR head:codex/ OR head:cursor/ OR head:copilot/), read from the count GitHub's search API returns.
  • It undercounts. Private repositories are invisible, and so is agent work that carries no signature or runs under a bot's own account. It can also count a person who names a branch codex/… by hand.
  • Most of these pull requests are opened in the author's own repository. A merged pull request there says little about quality; the board shows volume, not a ranking of agents.
  • Your recap and a repository's report are fetched by your browser from GitHub's public API, which allows ten searches a minute per network. We store a count of recaps made, not who was looked up.
  • If you claim your pass, GitHub signs you in with public information only. Your numbers are then counted once with your own sign-in and stored with your username, so your pass, the board and the badge load without asking GitHub again. The access token is deleted when the count is done.