What's your experience with AI-based code review tools?

[-]

Street-Remote-1004@reddit

I've been using LiveReview, Code Rabbi.

First, I solve the issues on my end, raise a new PR.

I generally get only logic comments from peers, no silly comments.

So its working pretty well.

[-]

Thats good, silly comments are probably the stuff I had to deal with most before haha. Haven't tried live review before, been using Codoki recently which has been a savior for me and my specific squad

[-]

divson1319@reddit

I’ve been messing around with AI code review tools for a while now… most of the early ones felt like a slightly smarter linter. They’d pick up style issues, missing docstrings, small bugs… useful for tidying up before a human review but not much beyond that. You still had to wade through the noise yourself.

When we started using CodeAnt.ai on our pull requests it felt different. It still flags the trivial stuff but because it pulls in context from the repo and history the comments are usually about why something is risky or inconsistent… not just what rule was broken. After a couple of weeks of feedback the noise drops a lot and the human reviewers can spend their time on design and edge cases instead of semicolons. It hasn’t replaced human review but it’s made our review cycles shorter and the feedback more focused…

[-]

PragmaticBoredom@reddit

The proper place to introduce AI feedback is before the PR goes up for review.

Have the original code author review any AI suggestions and act on them.

Forcing everyone to review the code and also review the AI suggestions wastes a lot of time.

Worst case, the team starts letting the AI code review substitute for their own thinking, or they spend all of their energy chasing what the AI tried to introduce into the conversation.

[-]

RozzSanz@reddit

Late to the party, but that's exactly what we are doing with codesync.com - let us know what you think. Simple plugin.

[-]

stellarcitizen@reddit

"introduce AI feedback is before the PR goes up for review" - Love that thought. I'm working on a tool that does exactly that so you don't pollute your PR with AI comments https://arcane.engineer/studio/pull-request-manager/

[-]

Sea-Confidence-1440@reddit

I used CodeANT, CodeRabbit and Atlassian CodeReviewer.

CodeRabbit is a winner for me :)

[-]

Alternative-Food-372@reddit

codeant for me, crazy intent aware results.

[-]

thewritingwallah@reddit (OP)

why CodeRabbit any good review examples?

[-]

Sea-Confidence-1440@reddit

Here is the example that I found very interesting:

120-132: State flags might desynchronise Magento & NetSuite

When we skip CM creation the code sets setCreditMemoCreated(!$order->canCreditmemo()).
Imagine a partial refund where canCreditmemo() is still true; the flag remains false, yet NetSuite will never accept a CM until the SO is fully billed.
Once the SO becomes billed, this worker won’t attempt CM again because the refund already exists in Magento. Consider persisting a “pending CM (wait for billing)” flag or re-queuing the job instead of marking it finished.

[-]

Grand-Ad3626@reddit

You might want to look at Gito

https://github.com/Nayjest/Gito

. It goes beyond basic linting — besides style it flags refactoring opportunities, performance and accessibility issues, and even security concerns. The usual setup is to run it on PRs before human review so trivial stuff gets cleared out and reviewers can focus on higher-level design and logic. It won’t replace humans, but it does consistently raise code quality and speed up the process.

[-]

Ordinary_Owl8281@reddit

I’ve experimented with tools like Cubic dev on my team. It’s been interesting, PRs get instant inline feedback and one click fixes, so you can catch small issues before a human review. Feels more like a productivity booster than a replacement for thoughtful review.

[-]

PapaDingDing@reddit

I treat (Cubic.dev) like an advanced linter. It learns from team conventions and enforces custom rules. In my experience, it speeds up PR workflow without reducing the value of human insight.

[-]

FeastyBoi23@reddit

I’ve used Qodo for code reviews, and it’s definitely more than just a linter. It helps catch bigger issues, like inconsistencies across files and missing best practices. The AI actually learns from past PRs, so it suggests things that are relevant to the way I work.

[-]

SidLais351@reddit

I’ve tried a few AI code review tools (Codacy, Codeium’s review feature, and recently Qodo for test-focused reviews). My take: they’re solid for catching consistently overlooked issues (missing null checks, test gaps, some anti-patterns) and enforcing style/consistency before human review, but they don’t replace a thoughtful human review for architecture or design discussions.

Think of them as a supercharged linter + test safety net. Qodo, for example, can auto-generate missing tests on PRs, which helps catch regressions you didn’t explicitly test and code review for, while Codacy flags style and potential bugs.

[-]

taluyev@reddit

The set of Apache-2.0 licensed projects for Rule Based AI Code Review.

The tools run on-premises with OpenAI-compatible LLM providers.

https://github.com/QuasarByte/llm-code-review-maven-plugin

https://github.com/QuasarByte/llm-code-review-cli

https://github.com/QuasarByte/llm-codereview-sdk

[-]

Significant_Rate_647@reddit

In my personal experience, I really do find AI code reviews leading to higher quality code. I mostly use Bito.ai.

Have a look at some recent pull requests I ran Bito's AI Code Review Agent on -

👉🏻 https://github.com/BitoGTM/Expense-Tracker/pull/12

👉🏻 https://github.com/BitoGTM/Expense-Tracker/pull/11

I've also used GitHub Copilot and CodeRabbit for ai code reviews, and also dropped a few blog posts comparing these tools. Read them here:

👉🏻 Bito vs CodeRabbit: https://bito.ai/blog/bito-vs-coderabbit/

👉🏻 Bito vs GitHub Copilot: https://bito.ai/compare/bito-vs-github-copilot/

[-]

Wolf171001@reddit

I’ve found that AI-based code review tools really work well when integrated right into your development workflow, especially for typed languages like TypeScript. Using Qodo’s VS Code plugin during form validation work has helped me catch type mismatches and subtle state bugs before I even run the app. It feels more like having a smart teammate who understands your entire codebase and spots issues that traditional linters might miss.

Qodo doesn’t just surface syntax errors; it analyzes your code context across multiple files and suggests fixes or improvements based on that. This reduces the load during PR reviews, letting the team focus on bigger-picture concerns like design and performance metrics like latency or responsiveness. From my experience, tools like this really help improve code quality and speed up the review process without adding noise.

[-]

kininkar@reddit

The amount of denial here is astronomical

[-]

earonesty@reddit

I wrote my own AI reviewer (coderev.q32.com)

I use it to review stuff for clients and for companies.

It grabs quite a large amount of relevant context before it does the review unlike some reviews

It also submits everything as draft with inline comments

The final result allows you to add your own voice call your attention to things and submit a real review

I don't think AI code reviews are good standalone right now although they are getting better

I updated the model recently to 4.1 or o3 and there was a substantial improvement

o3 has a hard time with line number locations , but it reasons more about causes.

anthropic is actually terrible and I'm going to remove it from the model selection list

I think it's excellent for assisting but I don't think it's great all by itself

[-]

cachemonet0x0cf6619@reddit

i don’t think the author should also be the reviewer /s

[-]

stackmatix@reddit

CodeRabbit is pretty good

[-]

aravindputrevu@reddit

We can use these AI reviewers as unbiased, org-set-standard-based reviewers. I see (while I have a conflict of interest) value in AI Coding Tools.

It all improves our coding styles and lets us write efficient code. While at large, I agree you need to make good have to be a reviewer of code generated using coding copilots.

[-]

raisingmonk@reddit

Good one, but here is my take with AI, you can play both the roles and complement each other

[-]

4lteredState@reddit

I want to return in a few hours to see "edit: removed '/s" on this post

[-]

Electrical-Top-5510@reddit

Claude to write and GPT to review

[-]

Snoo-82132@reddit

😂😂😂

[-]

terrorTrain@reddit

I used the trial of code rabbit and it pointed out some SQL bugs I was rather surprised it found.

But it isn't enough to justify the hefty price tags for them

[-]

bogdanstanga@reddit

You can try https://github.com/presubmit/ai-reviewer which is an open-source alternative to coderabbit. You would only pay your LLM usage with your model of choice.

disclaimer: i'm the creator

[-]

terrorTrain@reddit

Nice! I will check it out

[-]

bogdanstanga@reddit

For anyone interested in an open-source AI reviewer that is not spamming pull requests with noisy comments, you can try https://github.com/presubmit/ai-reviewer

disclaimer: i'm made this in a free weekend :)

[-]

Ghi102@reddit

We tried them, they never helped with much. The valid reported issues were usually minor, the kind of thing that I personally let slip by if I noticed it as the only issue in a PR. Many "issues" were questionable at best.

[-]

Frisky-biscuit4@reddit

I've been using Greptile and it's caught some crazy issues that would have gotten pushed to prod, happy to share examples

[-]

thewritingwallah@reddit (OP)

In my experience, automated AI review can find and identify bad patterns or practices in code or popular SDKs, but our company's tech stack is highly distributed and there is a lot of business domain logic that the AI simply can't really predict. let's say - If the comment string says something and the code does not do that, that's caught fine. But it is not so good at catching domain related bugs.

[-]

Forsaken-Ad-2791@reddit

If you don’t like a wall of text for every code review, I’d suggest devlo.ai. It gives high priority review comments. Can also handle write code for you as well (resolving github issues, bugs, unit tests, do followups on reviews etc). Runs and tests the code as well. More value than just a code reviewer.

[-]

Puzzleheaded-Gene806@reddit

Great question! I’ve been exploring AI-based code review tools, and while they’re not a replacement for human reviews, they can definitely speed up the process and help catch common issues early.

I’ve been trying out PullSense, which acts more like a smart assistant for PR reviews. Instead of just being a linter, it analyzes PRs, adds contextual comments, and highlights potential bugs or inefficiencies—all before a human review happens.

From my experience:
✅ AI helps surface low-hanging issues (e.g., unused variables, bad patterns, security risks).
✅ It speeds up human reviews by removing repetitive feedback cycles.
✅ Customizability matters—being able to choose different AI models makes a difference in quality.

That said, AI won’t replace deep architectural discussions, but it’s super helpful in making PRs cleaner before human review. 🚀

Curious to hear others' experiences—what has worked (or not) for you? 🤔

[-]

thumbsdrivesmecrazy@reddit

Here is a good article below that highlights most popular code review tools available as well as how it can significantly enhance the development process: 14 Best Code Review Tools For 2025

[-]

pennythenomad@reddit

pullreview.ai is cool

[-]

rushout09@reddit

I have been using a code review tool codeant.ai that mixes ai and sast and works perfectly. I like their approach to AI vs traditional code review

[-]

thewritingwallah@reddit (OP)

codeant is good one but have some flaws compare to coderabbit - I compared them here -https://www.devtoolsacademy.com/blog/coderabbit-vs-others-ai-code-review-tools

[-]

PulpFunction412@reddit

Coming a little late to the convo but I can offer some insight OP...

AI code review tools can be a helpful supplement, especially for catching minor issues or maintaining consistency. While they’re not perfect, they help streamline the process before submitting for a human review. Concerning "code review being hard", in my opinion, it's asking a lot out of an engineering manager/lead who is reviewing code to 1) remember 100s of rules their organization is trying to enforce and 2) accurately & efficiently review the code based on that ruleset. Most aren't removing the human element of code review either, just speeding up the tedious and monotonous work that I can't imagine a lot of devs dream of doing, which is comparing the submitted code to see if it breaches any rules, best practices, or policies.

I’d recommend giving tools like Tabnine a try. While they won’t replace human reviews, they can reduce the grunt work and let you focus on higher-level improvements. IMO, Tabnine feels like a smarter linter—it’s great for catching repetitive mistakes, but I wouldn’t rely on it for anything requiring deep logic or architecture validation.

[-]

aravindputrevu@reddit

Disclaimer: I work at CodeRabbit

One thing that I like about AI Code Review tools -> they let me take the first pass at reviews. These AI Code Reviewers can catch interested issues, ofc based on the tool you use. For ex: CodeRabbit can do automated lining, bubble up non-noise lint suggestions and generate fixes. That's a win for me.

As the time passes, my grunt work is automated, I'm working on logic, which is why I'm a programmer.

[-]

iama_regularguy@reddit

I use GitHub copilot at work and I've never seen a useful suggestion. I almost never see suggestions in general. And a few suggestions were super wrong. Like using a hard coded string instead of an i18n key. Seems like it should've known all our strings are internationalized.

I treat it like a linter after all our other linters. Mainly it's just an experiment.

[-]

aravindputrevu@reddit

how about Cursor?

[-]

thumbsdrivesmecrazy@reddit

AI code review tools at some extent automate the review process by providing a structured layout, simplifying code access and understanding. Here is a quick guide introducing and comparing some top code review tools for 2024: Code Review Tools For 2024 - Guide

PR-Agent
GitHub
Bitbucket
Collaborator
Crucible
JetBrains Space
Gerrit
GitLab
RhodeCode

[-]

col-summers@reddit

I use obsidian with the text generator plug-in that integrates it neatly with an LLM. I create a giant prompt that includes the architecture document, any jira tickets, chat conversations, and finally the entire git diff. I can't report any specific details of hitting a home run in terms of discovering a bug I wouldn't have otherwise found but I feel like it has been useful though in sort of a vague and non-specific way. I can tell you specifically it's very useful for collecting all of my notes and then composing that into a comment I can drop into the code review. I guess what I'm saying is I use it to improve the quality of my written thoughts but LLM's really cannot on their own generate logical conclusions.

[-]

Nyphur@reddit

Interesting, could you give an example? I didn’t know obsidian had LLM integration

[-]

col-summers@reddit

Here's the link to the plugin

https://github.com/nhaouari/obsidian-textgenerator-plugin

I can't really give an example because what I'm describing is what I do for work. But it's really just like I said: copy and paste all the content from all the documentation, chat messages, code etc into a giant prompt and use that as your work surface for interacting with the AI about the project.

[-]

Orca-@reddit

I’ve used the copilot review and I wish we would turn it off. It generates work and rarely catches something worth caring about even a little.

Trash, I would pay money to not have it shit my PRs up.

[-]

mugwhyrt@reddit

Does your work require you to use copilot? I'm confused why you can't just turn it off or at least ignore its suggestions

[-]

Orca-@reddit

I can ignore the suggestions but it’s default on for PR feedback. I haven’t felt like incurring the political hit for pushing back on that.

[-]

ezaquarii_com@reddit

It started cargo cult war. Incompetent contributor with AI support is still incompetent, but also pretentious.

- This brilliant AI tool says you are wrog

- My code works ok, I'm using X technique

- But the AI tool is smart and it says it's bad

- X has been known for Y years now and is widely respected practice

- But the AI tool is smart and it says it's bad

[-]

bloudraak@reddit

I use every damn tool at my disposal; it’s never this vs that (unless there’s some legitimate reason; often legal, compliance etc).

I find AI to be at a level of most junior folks, even surpassing some senior folks. It’s often better than Google search (since that’s often incomplete opinions that’s often useless).

For code reviews, I find it to be useful, since it often finds bugs the linter etc doesn’t find. Better yet, it can explain many defects, which many tools lack. But often it comes up empty — that doesn’t mean I should throw it out — most folks come up empty during code reviews.

But like working with juniors, search results, tool outputs, I’ll vet everything I learn.

Many complain about AI being useless; but in my experience, I’ve been able to develop software in a matter of hours, that would have otherwise taken my days or even weeks to write. And it was good enough.

Don’t let perfect be the enemy of the good.

[-]

svvnguy@reddit

I evaluated current LLMs for this purpose and they're all quite terrible. They will miss the point and even if you explain it to them, they'll say OK and then proceed to output garbage again, because there are nuances that they haven't encountered in their training.

On top of this, they code at early junior level, so not something I would rely on for any sort of review, unless you want to waste your time.

[-]

Wonderful-Habit-139@reddit

Pretty sure they code at a level worse than juniors.

[-]

tetryds@reddit

Pretty sure juniors code at their level

[-]

Wonderful-Habit-139@reddit

Do you think chatgpt or claude is better than for example someone that coded neetcode io? Because that was a junior.

[-]

tetryds@reddit

Technical ability does not translate directly to seniorship.

I meant that juniors are using these tools so much that it's hard to distinguish between the two.

[-]

Wonderful-Habit-139@reddit

Objection! Relevance?
I don't know who said technical ability translates to seniorship.

That aside, it's a shame when juniors rely too much on these tools.. The other day I saw a post of a junior that didn't even know basic syntax and used chatgpt for everything. I don't think that person is even junior level imo, and shouldn't have been hired.

But juniors that are actually legitimate and get accepted in FAANG or at least a company that has a decent interview process, are definitely better than LLMs.

[-]

runmymouth@reddit

I fine cve vulnerability tools and a good linter to be far more efficient than ai.

[-]

jek39@reddit

0

[-]

koreth@reddit

We started using one of them (Graphite Reviewer) on my team about a month ago and it's been mixed.

On our backend code base, where I spend most of my time, it's been close to useless. It has caught a handful of legitimate minor things but I'd say at least 75% of its comments are off base in some way. Often not wildly wrong, just lacking context. An example from earlier today was it not knowing that Flyway runs SQL migration scripts in transactions, so it commented that two of the statements in one of my scripts needed to be wrapped in a transaction. That's when it comments at all, which it rarely does.

On our web frontend code base, which uses React, the devs find it very helpful. It points out stuff like, "The properties you're putting on this child component actually need to be on the parent component, and here's why" and from what I can tell (as a mostly backend-focused guy) it's usually right. It has also apparently helped a lot with fixing mistakes in our usage of third-party libraries.

I think a big part of the difference is that it has been trained on way more React code examples than enterprise backend monolith code examples.

Another difference: on our team, the frontend devs are much more junior.

I'm honestly not sure which of those two factors is larger.

[-]

thewritingwallah@reddit (OP)

most AI code review tools were likely trained on popular open source projects and common patterns

give Coderabbit https://www.coderabbit.ai/ a try. It might handle backend code better since it's trained specifically for deeper code analysis and I quickly checked Graphite it's pricing is very high compare to CR.

CR pricing is more developer-friendly:

Pro tier at $12/month vs Graphite's $99/month
Coderabbit is free forever for open source projects
No seat limitations

[-]

Connect-Clock-9778@reddit

Probably the latter. We tried similar tools and it doesn't really catch anything other than questionable style issues or the occasional missing key. There just aren't that many issues to catch in most of our PRs.

No one on our team is a junior developer. The few times we've had a primarily backend dev open a PR it didn't catch anything we wouldn't have said and missed quite a few subtle business logic issues.

Putting the props in the right place is like year one of React. If AI helps catch those then more power to them but I'd be asking why it happened in the first place.

We ended up disabling it on our codebase.

[-]

tristanbrotherton@reddit

https://codepeer.com/blog/ai-sucks-at-code-reviews

[-]

UltimateTrattles@reddit

I have yet to see one that adds more signal than noise. But I believe one will eventually arrive.

I even use ai every day to code - but even that isn’t as good as advertised and requires you to be pretty knowledgeable already to get value from it.

[-]

Dry_Ruin_1743@reddit

a good engineer leaving detailed and contextual feedback isn't even in the stratosphere of LLM capabilities

[-]

Perfect-Ad4901@reddit

I use GitHub CoPilot at times with PR reviews. In the beginning, I found value with it, being able to ask questions or offer suggestions. I don't think I've used it in a while.

Not directly answering your question, but I use ChatGPT a lot in my coding now. It moves me forward so quickly and helps to distill complex things into something simpler. It's a huge unlock and far, far from perfect, but damn is a great tool.

[-]

rasplight@reddit

I've built AI review comments into my own code review app at codelantis.com (you can have the AI review the current file from the PR diff). In my experience, this can be very helpful to get a first (and fast!) review.

What fascinates me is that it sometimes seems like the AI (GPT-4o) has good and bad days, too. Sometimes it will point out issues immediately, and sometimes it will slip them through 😅