Patternis

Pull-request review that routes work by fit and keeps receipts you can read.

Your provider key never reaches us. Patternis dispatches work into your GitHub Actions runner, where it uses a key held as a secret in your repository. We receive a receipt. That is the whole architecture, and it is the reason setup asks you to do two things yourself.

What it does

On a pull request, it picks a model tier by what the task needs and how much headroom your provider has left, runs the review on your runner, puts the review on that run's summary page, and posts a receipt on the pull request that links to it. Every decision, refusal and result gets a receipt. A workflow in a repository can read that repository's receipts, including when and why Patternis declined to act.

See an example review, with its source limits, setup steps and separate AI costs.

What it asks for

`metadata`	read — mandatory for every GitHub App
`pull_requests`	write — to receive the events and post the receipt on the pull request, which GitHub allows only with this
`issues`	write — to post the receipt where the work happened
`actions`	write — to dispatch into your own runner
`members`	read, organizations only — to check that the person subscribing is an owner

It cannot modify your code. contents is not requested at all: the runner reads your diff with your own Actions token, not with the app's.

What it costs

15 days free, no card. Then $20 a month plus tax where it applies. Your AI provider bills the AI usage to your own key, on your own rate limits. The pricing page explains what a unit of work is and what the trial includes.

Setting it up

Installing takes one click; it will not do anything until you also add the workflow file and your provider key. The page you land on after installing walks through both. The workflow is a readable download before you install anything — it is the code that will run on your infrastructure with your key, and you should read it.