I Switched to an Open Source Code Reviewer
And I'm Not Looking Back

Be honest.. how much of your code are you actually reading?
Not the quick glance before merging.
I mean going through the changes, checking the logic, thinking about security, making sure the thing your agent just wrote makes sense.
I’m asking because my previous attempt at automated review was an OpenCode GitHub Action with a prompt that basically said, “review the changes in this pull request thoroughly.” (I actually made it worse by using a free model 🤣) Well done mate. You nailed it. Software engineering solved.

It worked as an experiment, but I wanted something built for the job.
That’s where PR-Agent comes in, and why I’m switching my personal projects over to it.

​
This isn’t a CodeRabbit breakup

Let’s get this out of the way: CodeRabbit is an amazing product, and I’m still a happy paying customer at work.
This isn’t me discovering that commercial software is evil.
For my own projects, though, I wanted an open-source reviewer I could configure and run myself, without another per-developer review subscription.

And to be fair, CodeRabbit already offers free reviews for public repositories.
So “open source means free reviews” isn’t the whole argument here.

For me, it’s control.
My workflow, my configuration, my choice of model.
Including when the repository is private.

PR Agent is now community maintained under its own GitHub organization, following Qodo’s donation of the project.
This just makes a stronger argument for me - literally a project customers paid for, made open source for everyone to use.

​
Better than “please review make no mistakes”

The difference isn’t just a longer prompt (I know that's what you think..)
PR-Agent has tools for reviewing changes, describing pull requests, answering questions, and suggesting improvements. It also deals with the less exciting work around those requests: preparing the diff, managing the available context, and fitting it into the model’s token budget.

That last part matters when an agent casually hands you a PR the size of a small novel.

Its compression strategy filters and trims what gets sent when the change is too large.
That doesn’t mean the model magically understands everything, it means there’s an actual strategy beyond throwing the whole thing into a prompt and hoping it works. Which, honestly, is already an improvement over my previous setup (low bars...)
​

Put it where the review already happens

I went with GitHub Actions.
No separate review dashboard to keep open.
The feedback shows up on the pull request, where I’m already deciding whether to merge it.

The GitHub setup needs a workflow under .github/workflows/ and credentials for the model provider.
In this setup, that’s an OpenAI API key, added as a repository secret:

GitHub supplies the workflow’s GITHUB_TOKEN (did you know that?! -> if you have something like:

The token will be set for you by the platform(!)
Merge the installation into the default branch, open the next PR, and the configured workflow can review it automatically.

You can also leave commands directly in the PR discussion. Request a review:

Ask for a description:

Or request code suggestions:

That’s the interaction I wanted -not copying a diff between windows, the reviewer is part of the PR.

​
Don't settle for a "raw" review

PR-Agent supports committable suggestions!
You can enable them in the repository configuration, .pr_agent.toml, instead of asking for the same behavior manually every time. Set:

Give it something better to work with

The configuration goes beyond picking a model.
You can add custom review instructions, choose fallback models, and control token limits so it doesn't go wild.
There’s also support for skills, so you can give it guidance relevant to the code it’s reviewing instead of expecting generic advice to cover everything.

Another useful piece is dynamic context.: A diff alone doesn’t always explain what a function is doing.
PR-Agent can expand the surrounding code to give the model more context, without simply sending every entire file.

The configuration reference is worth looking through.
These are the kinds of controls that made me want a purpose-built reviewer in the first place.
​

Open source doesn’t pay the API bill

There isn’t a PR-Agent seat subscription in this setup. There is still model usage, and potentially GitHub Actions usage, to pay for.

Also, running the reviewer yourself doesn’t automatically make inference private.
With my OpenAI backed setup, review context goes to openai API.
A private GitHub repository doesn’t change that.
Only use code you’re authorized to send to that provider.

So no, this isn’t “everything is free and nothing leaves your machine.” It’s a different tradeoff, with more of the setup and responsibility on you.

For my personal projects, that’s a tradeoff I can work with.

I’m not claiming this thing beats every commercial reviewer, or that another AI checking your AI means you can stop reading code (not opening that debate!)
I’m saying it’s a welcome upgrade from my homemade “review this thoroughly” prompt, and it gives me a review workflow I can actually shape, for [kinda] free (using exiting subscriptions).

Start with one repo. See whether the findings are useful, check the suggestions, run your tests, and look at the usage bill.

The goal isn’t more comments on your PR, It’s catching something worth fixing before you merge.

I hope this was valuable!
Thank you for reading.
Feel free to reply directly with any question or feedback.
Have a great weekend!