Cowboy Docs Evidence Get Cowboy free

Sort what’s coming. Check what happened.

Cowboy sorts every new pull request as it arrives, so you know where to look first. It looks back at the ones that merged, to see what went wrong and whether the sorting, or any rule of yours, would have caught it. And it shows which of your team's habits pay off in your repo, then turns all of that into instructions your coding agents follow. The screens here come from the demo that ships with Cowboy, which runs on a made-up repo, except where a caption says otherwise. All of it is free. This is a preview release: the last section says what's tested and what isn't yet.

Sorting, as pull requests arrive.

Every new pull request lands in easy, normal or careful by fixed rules, with the reasons written on it, and agent work is tagged. Treat the pens as a starting point: the next part shows how Cowboy checks them on your repo.

The corral: every open pull request, live.

One page shows every open pull request, sorted into three pens, with agent work tagged. It updates itself as Cowboy works.

  1. 1LiveNew pull requests drop into their pen the moment they're sorted. No refresh.
  2. 2Three pensEasy: probably fine to skim. Normal: a regular review. Careful: worth a closer look.
  3. 3Doubles and straysTwo agents on the same issue, or an agent that went quiet after you asked for changes.
  4. 4Off task?With the optional local model, Cowboy marks changes the task never asked for. In preview.

Every pull request, explained.

Click any card to open its page. Everything Cowboy knows about that pull request is in one place.

  1. 1Why this penEvery rule that fired, in plain words. Same pull request, same pen, every time.
  2. 2What Cowboy measuredSize, how many files depend on the change, sensitive paths, shortcuts and checks.
  3. 3Local model notesA short brief and where to look first, from an open-weight model on your server. Optional, in preview.
  4. 4Your callDisagree? Move it to another pen. Cowboy updates the label and counts it in the tally.

Works where your PRs already live.

On GitHub, Cowboy only adds labels and one comment, edited in place on every push, and asks the code owners of careful pull requests for a review. GitLab (gitlab.com or your own), Forgejo and Gitea work the same way and are in preview.

  • On GitHub, connect with one click: Cowboy creates its own private GitHub App, and you pick the repos.
  • It never merges, closes or reviews code.
  • Change the label yourself and Cowboy remembers.

Rules you can read.

Sensible defaults out of the box. Change them per repo with one small file on your default branch. A pull request can't loosen its own rules.

  • Pick your own sensitive paths.
  • Name the bot accounts your agents use.
  • Turn the experimental check for unmarked pull requests off.

Looking back, after they merge.

This is the part that earns trust. Cowboy follows every merged pull request, traces trouble back to the one that caused it, and checks whether the pens, or simpler rules, would have pointed at it. In our own tests the careful pen didn't beat simpler rules, and Cowboy will tell you the same if it's true on your repo.

Trace trouble to the pull request.

Cowboy follows every merged pull request for 30 days and ties what happened back to it: reverted, broke main, needed a fix, or caused an incident.

  • Reverts and a broken main count by themselves. Follow-up fixes and incidents from Sentry, PagerDuty or any tool are suggestions a person confirms.
  • Don't wait a month: rate the last 90 days of merged pull requests now. Nothing is written to your git host.
  • A track record for every agent, by area of the code, and a report card for every rule.
  • A tamper-evident record of every sort, move, approval and outcome, ready to export.

Test your rules against simple alternatives.

A rule that flags a quarter of your pull requests will catch some trouble by luck. So Cowboy gives the same share of pull requests to simple alternatives and compares.

  • Your careful pen against “largest first”, “most files touched first” and a random pick. Your easy pen against the same share of smallest pull requests.
  • Counts and a likely range for every result, in words: outperformed, no demonstrated advantage, did worse.
  • Fewer than 20 pull requests with trouble in a repo? It says “Not enough evidence yet” and stops there.
  • The same calculation as the published evidence, on the After merge page or with cowboy evaluate.

Real numbers, not the demo: Cowboy's own rules, replayed on next.js, Airflow and Kibana history.

Your team's habits, judged on your history.

How often pull requests here change tests with code, avoid shortcuts, stay a reviewable size, get reviewed and merge with checks green. And whether each habit went with less trouble after merge, in this repo.

  • A share for each habit over the last 90 days, next to the 90 days before.
  • The same verdicts as the rest of Cowboy: holds here, did worse here, no demonstrated difference, not enough evidence yet. Size is held fixed.
  • When a habit your team requires shows no difference and the range is tight, it says so: that rule may be a ritual worth dropping.
  • Numbers for the whole team, never per person.
  • Why per repo: across nine public repos, no habit went with fewer reverts everywhere.

Real numbers from the next.js history in the demo. “Did worse here” is a question to look into, not a reason to write fewer tests.

Guide your agents.

Coding agents follow the instructions file in your repo. Cowboy writes a section for it from what it already knows, so agents open pull requests your team can review quickly.

  • Which paths and files get a careful review, and the size that gets the quickest one.
  • The shortcuts Cowboy flags, only for the checks you have switched on.
  • This team's habits, with evidence quoted only where your history backs it.
  • How to say an agent helped, so reviewers know what they're reading.
  • Copy or download it from the Habits page, or run cowboy guide --write AGENTS.md. It replaces only its own section, between markers. Cowboy never commits it for you.

Shortened. The full section lists every careful path and shortcut check.

The weekly tally.

Which agents send the most careful work, cut corners, or leave strays behind. And how often your team moved a pull request to another pen.

  • A page in the dashboard, posted to Slack or sent by email every Monday (you pick the day and hour).
  • Counts only, never code.

Missed webhooks get caught.

Webhooks get lost. Every 5 minutes a catch-up sweep finds any pull request Cowboy missed, and the Activity page lists everything it did, newest first.

  • When nothing changed, it costs about one API call per repo.
  • Human pull requests are checked once and left alone.

Set up in a minute, in your browser.

Run cowboy and your browser opens on a setup page. No config files, no environment variables.

1

Connect GitHub.

One click if you use GitHub’s command-line tool. Otherwise the page opens GitHub’s token form already filled in with read-only access: tick your repos, generate, paste.

2

Pick your repos.

Choose from the repos your login can see. Cowboy sorts their open pull requests and rates the last 90 days of merged ones.

3

Start by just looking.

Nothing is written to GitHub until you switch labelling on, and Cowboy checks it has permission before it does.

■ Runs on your machine■ Your token stays on it, readable only by you■ Remembered next time: just run cowboy■ Team server? A GitHub App, set up from the same page
preview

What's in the preview.

Cowboy is a preview release. Here's what has been tested on real code, what works but is still in preview, and what isn't built yet.

Tested on real code

  • The sorting rules, tested on about 28,000 merged pull requests from eleven public repos: the easy pen mostly held up; the careful pen didn't beat simpler rules. See the evidence.
  • The comparison against simple alternatives: it reproduces the published evidence exactly.
  • Labels, the one comment, doubles, moving a pull request and noticing closes, live on a real GitHub repo.
  • The catch-up sweep and the dashboard login.
  • Blast radius (how many files use a changed file), checked against the TypeScript compiler and Python's own parser on Sentry, Grafana and n8n. It downloads the repo from your git host itself, so no local copy is needed.

Works, in preview

  • Setup in the browser: run cowboy, connect with your GitHub CLI login or a token, pick repos, and switch labelling on when ready. Tested against a stand-in for GitHub, not yet github.com.
  • The GitHub App route for a team server: tested against a stand-in for GitHub, not yet github.com.
  • GitLab, on gitlab.com or your own server: tested against a stand-in for GitLab's API, not yet gitlab.com.
  • Forgejo and Gitea.
  • Storage in SQLite, saved change by change. An older install's data file is moved over by itself on first start.
  • The Habits page and the agent guide: covered by tests and run on the demo's next.js history. Reviews and checks need a live install to fill in.
  • Local model notes and the "off task?" flag.
  • Review requests to code owners, and strays: covered by tests, not yet seen on a live repo.
  • The experimental check for pull requests with no agent mark.
  • The weekly tally in Slack or email: covered by tests, not yet sent to a real workspace.
  • After-merge tracking: revert detection checked on two years of Sentry, Grafana and n8n history (it traced 70–99% of reverts to their PR); the rest tested against stand-ins for GitHub, Sentry and PagerDuty.

Not built yet

  • Review requests to teams like @org/team. Today it's individual people.
  • A signed Mac app.
  • Rules suggested from your own history, for a person to approve. The fitting is built and tested, but it didn't beat chance on fresh repos, so it isn't switched on.

Free. Open source. Your server.

All of it. Whatever's free today stays free.

Get Cowboy free Read the docs Back to home