Subscribe
Sign in
Graphite Reviewer — Your high-signal AI code review companion
An always-online AI teammate who gives you consistent, high-signal, and actionable feedback on every pull request in seconds. Graphite Reviewer catches bugs, enforces consistency, and gives you back time to build. Now available to teams on GitHub.
47
Replies
Love the approach with Graphite where the tool is truely an augment to the workflow of PR code review. It's really just another pair of eyes for the code. Also it's great to see it's customized to the codebase, and set follow custom rules - lots of flexibility and personalization there.
Congrats on the launch @merrill_lutsky and team!
Hi @sentry_co! I haven't tried Cursor's PR review personally, though I think their product is fantastic, and the team behind it is stellar. From what I understand, Cursor is more focused on code creation, whereas Graphite is more focused on leveling up the review and collaboration side of changes.
One perk of Graphite Reviewer is that it works automatically across all PRs in a repo, regardless of how they were created. So, even if one teammate uses Cursor, another Vim, and another XCode, everyone can benefit from the comments of a smart Graphite Reviewer on GitHub. I want developers to be free to pick their favorite tools while helping ensure a high-quality bar for every change merged.
@greg_foster Yeah its beta. But it's fantastic. You can try it out by toggling it on in the prefs of cursor. the thing with cursor is that it's also becoming a research assistant (also beta). it goes out and finds 5-6 blog posts about your prompt and then uses that as its research before giving you its answers. Which makes the answers far better than other copilots. But it's more like a review before you make your review to the team. So not directly competing with online AI reviewers.
Congrats to the Graphite Reviewer team on the launch! This AI teammate sounds like a game-changer for speeding up the code review process while ensuring quality and consistency. How customizable is the feedback for different coding standards or team preferences?
@vietpham Reviewer is aware of your codebase out of the box and will learn your team's patterns and conventions! We're also shipping custom AI prompts and regex matching rules so you can further customize Reviewer for your team.
The first time I used this product I was so blown away I shared this South Park clip with all my friends:
Love the pricing/plan.
Helps a budding developer, upcoming startup ecosystem to work and have flawless code base.
In practicality, I guess a developer could provide better insight on having LLM insight.
Excited about more innovation in this particular domain and affordability.
Good luck to them team and congratulation on the launch!
@evan_paul I appreciate the kind words! In a world where more and more code is being created (and written by LLMs), I think it's critical we better empower reviews. We _need_ humans in the loop to ensure what gets merged in is safe and right, and I think giving those reviewers a second pair of eyes will help them do their jobs better.
Wow! We spend so much time on reviews, and usually the errors we found are the style one. But static analyzers does no satisfy us in many cases. It looks like a savior!
@aleksanadr_lavrinenko Agreed! Static linters are great for their deterministic brutality - but they're also brutal, and it's hard to write new ones. Graphite Reviewer adds a wonderful generic catchall and is smart enough to consider older merged code changes as an implicate style guide in your codebase. The goal is to get authors to feedback faster and let reviewers focus more on high-level architecture rather than minor nits and semantic bugs.
Congrats on the launch @merrill_lutsky & team! This looks like a game-changer. Wishing you all the best as you grow 🚀🙌
Thanks @tomas_blatak1! We've been loving using Graphite Reviewer internally, and the feedback from our private beta has been wonderfully positive. Give it a spin and let us know what you think!
Hi Merrill! Graphite Reviewer sounds like a game-changer for code reviews. The focus on eliminating noise and providing actionable feedback is fantastic. How does the tool handle different programming languages?
@web3_antivirus it handles all programming languages out of the box - this is inherited from the base models we're using! Appreciate the kind words - excited for you to give it a try!
@krishyadav007 Yup! You can use Graphites' local stacking functionality through our VSCode extension and CLI. The Graphite Reviewer experience happens in the cloud whenever you create a pull request in your repository. We dont yet support running Reviewer locally, but it's something we're considering for the future :)
I'll be honest -- I was skeptical that an LLM could come anywhere close to a human's ability to review code accurately. My experience with Graphite Reviewer has been incredible though. Its comments have been useful and actionable nearly every time, and it spots issues that I don't think I would have noticed otherwise. Very cool to see the skeptics (like myself!) be convinced that AI code review is here to stay.
This sounds like a game changer, Merrill! I'm really intrigued by the <3% false-positive rate you mentioned. That kind of precision is what a lot of us have been looking for in AI tools. Can you dive a bit deeper into how Graphite Reviewer leverages RAG to ensure context-aware feedback? Also, will there be integrations with CI/CD pipelines in the future to automate testing after the code review?
Lastly, since it's free for October, is there a planned pricing model for after this period? Really excited to see how this can boost team productivity! 🚀
Graphite reviewer is amazing! We've been using it internally on our team and it's been a huge uplevel.
I can't decide what I appreciate more: the markdown formatting in the comments, or the sweet sweet jams in the promo video. Congrats on the release, this looks amazingly useful! Particularly on my solo projects where I tend to be far too lenient in my code reviews...
Hey Merrill,
I'm particularly interested in how you've achieved such a low false positive rate. What techniques are you using to filter out hallucinations and noise?
How does the codebase-specific RAG work? Does it require any setup from the user?
Congrats on the launch!
Hi @kyrylosilin! Great questions!
> What techniques are you using to filter out hallucinations and noise?
The main things that have helped us get a low false positive rate have been
1) Building a detailed eval battery to test any changes we make to prompts
2) Leveraging the most expensive frontier models
3) Feeding in exceptionally large prompts with as much context as possible, taking advantage of recent increases in token limits
4) Creating long pipelines of LLM queries, often 5+ queries chained together to refine and evaluate anything outputted by the LLMs to ensure usefulness
5) Comparing existing code changes with older code and code changes in the repo. Most reviewers are subjective, but the real truth lies in what already exists on main
> How does the codebase-specific RAG work? Does it require any setup from the user?
No setup is required! After enabling Graphite Reviewer, our servers will asynchronously index your existing codebase in the background and leverage the indexed code in future reviews. It happens seamlessly behind the scenes, but Graphite Reviewer will cleanly cite any references so that users know where the context is coming from.
Graphite