You need an agent, somewhere to keep files, and a way to reach your pull requests. All three are already yours. What follows installs the two skills we use for the drafting - any skill that writes the same file works as well - and after that, reviews arrive on their own and wait for you. Nothing reaches GitHub until you send it.
One address carries both skills and the queue helper they need, inline, so there is nothing to chase down afterwards. Hand it to your agent:
read https://dev.review/install and follow it
Or fetch it and do it yourself. It is a document rather than a script: every block in it is one file, with the path it belongs at written above it, so you can read the whole thing before anything lands on your disk.
Name a drafts directory in your CLAUDE.md, and attach a source in the app to that same directory. That is the whole integration: the sweep writes there, the app reads there, and nothing sits between them.
A folder on this computer works, and so does a bucket you own, a GitHub repository, or any git remote - none of them ours, which is why there is no account to make here.
Run the sweep. It finds the pull requests waiting on you, drafts each one, and stops. Open the app and the reviews are there, with the clean lenses shown beside the flagged ones and the QA evidence underneath.
Then you do the part that is yours: cut what is wrong and send what is left, under your own name.
Each skill is one markdown file in its own directory, and the sweep's queue helper sits beside its SKILL.md in a collector/ directory:
~/.claude/skills/dev-review/SKILL.md ~/.claude/skills/dev-review-sweep/SKILL.md ~/.claude/skills/dev-review-sweep/collector/
Running the app yourself, and what the GitHub token needs to be scoped to, are both in the documentation.
dev-reviewDrafts one pull request. It claims the file first, so two runs cannot fight over the same review, then rewrites it as each lens finishes, with a note on where it has got to. The app shows all of that as work in progress and keeps posting locked until the run sets finishedAt.
Unattended by design: it asks nothing and waits for nothing, because the person it is drafting for is not at the keyboard.
dev-review-sweepFinds the pull requests awaiting your review, including your own, and drafts the ones that have no draft yet. It hands the reviewing skill its unattended mode, which means the QA runs without asking rather than not at all. One failure is reported and the sweep carries on to the next.
It carries a small queue helper, which is the part that decides what is new and what got deferred.
Both files are inside the install document, whole, each under the path it belongs at. It is generated from the skills themselves, so reading it is reading what lands.
Neither of these is privileged. A skill is a thing that writes one JSON file into storage you own, and the app reads whatever it finds there. The schema is published and versioned, so a skill written against it is a skill this interface can read, whoever wrote it and whatever agent runs it.
The whole integration is a file on disk.
Which is also the honest reason these two are worth reading rather than trusting: they are one way of producing that file, and the format is the part that matters.