Graphite is joining Cursor(cursor.com) Graphite's announcement: https://graphite.com/blog/graphite-joins-cursor |
Graphite is joining Cursor(cursor.com) Graphite's announcement: https://graphite.com/blog/graphite-joins-cursor |
My other question is whether stacked PRs are the endpoint of presenting changes or a waypoint to a bigger vision? I can't get past the idea that presenting changes as diffs in filesystem order is suboptimal, rather than as stories of what changed and why. Almost like literate programming.
- Hunter @ Ellipsis
what does graphite have to do with code review?
Personally, I work on Graphite for two reasons. 1) I love working with kind, smart, intense teammates. I want to be surrounded by folks who I look up to and who energize me. 2) I want to build bleeding-edge dev tools that move the whole industry forward. I have so much respect for all y’all across the world, and nothing makes me happier than getting to create better tooling for y’all to engineer with. Graphite is very much the combination of these two passions: human collaboration and dev tools.
Joining Cursor accelerates both these goals. I get to work with the same team I love, a new bunch of wonderful people, and get to keep recruiting as fast as possible. I also get to keep shipping amazing code collaboration tooling to the industry - but now with more resourcing and expertise. We get to be more ambitious with our visions and timelines, and pull the future forward.
I wouldn’t do this if I didn’t think the Cursor team weren’t standup people with high character and kindness. I wouldn’t do this if I thought it meant compromising our vision of building a better generation of code collaboration tooling. I wouldn’t do it if I thought it wouldn’t be insanely fun and exciting. But it seems to be all those things, so we’re plunging forward with excitement and open hearts!
What do you like about the non-AI parts? I mean it's a little convenient to be able to type `gt submit` in order to create the remote branch and the PR in one step, but it doesn't feel like anything that an alias couldn't do.
With more resources than ever. We're building whole platform. That's a lot more than just AI.
Somebody screenshot this please. We are looking at comedy gold in the next 3 years and there’s no shortage of material.
As someone who is a huge IDE fan, I vastly prefer the experience from Codex CLI compared to having that built into my IDE, which I customize for my general purposes. The fact it's a fork of VSCode (or whatever) will make me never use it. I wonder if they bet wrong.
But that's just usability and preference. When the SOTA model makers give out tokens for substantially less than public API cost, how in the world is Cursor going to stay competitive? The moat just isn't there (in fact I would argue its non-existent)
Is it market share? Because I don't know who has a bigger user base that cursor.
> After bringing features of Supermaven to Cursor Tab, we now recommend any existing VS Code users to migrate to Cursor.
Supermaven was acquired by Cursor and sunset after 1 year.
The idea is to hook into Bitbucket PR webhooks so that whenever a PR is raised on any repo, Jenkins spins up an isolated job that acts as an automated code reviewer. That job would pull the base branch and the feature branch, compute the diff, and use that as input for an AI-based review step. The prompt would ask the reviewer to behave like a senior engineer or architect, follow common industry review standards, and return structured feedback - explicitly separating must-have issues from nice-to-have improvements.
The output would be generated as markdown and posted back to the PR, either as a comment or some attached artifact, so it’s visible alongside human review. The intent isn’t to replace human reviewers, but to catch obvious issues early and reduce review load.
What I’m unsure about is whether diff-only context is actually sufficient for meaningful reviews, or if this becomes misleading without deeper repo and architectural awareness. I’m also concerned about failure modes - for example, noisy or overconfident comments, review fatigue, or teams starting to trust automated feedback more than they should.
If you’ve tried something like this with Bitbucket/Jenkins, or think this is fundamentally a bad idea, I’d really like to hear why. I’m especially interested in practical lessons.
My team has been using Qodo for a while now and i've found it to be pretty helpful. EVery once in a while it finds a serious issue, but the most useful part from my experience are the features that are geared towards speeding up my review rather than replacing it. Things like effort labels that are automatically added to the pr and a generated walk through that takes you through all of the changed files.
Would love to see a detailed comparison of the different options. Is there some kind of benchmark for AI code review that compares tools?
Obviously the working tree should be a commit like any other! It just makes sense!
The problem however lies in who or what does this rebasing in a multi-tenant environment. You sort of need a system that can do it automatically, or one that gives you control over the process. For example, jj can often get tripped up with branch rules in git since you might accidentally move a bookmark that isn't yours to move, so to speak.
Huge fans of their work @ GitStart!
for anyone else looking for a replacement, git spice and jujutsu are both fantastic
Then Cursor takes on GitHub for the control of the repo.
Given the VP of GitHub recently posted a screenshot of their new stacked diff concept on X, I'd be amazed if Graphite folks (whos product is adding this function) didn't get wind of it and look for a quick sell.
So, we'll see what it ends up like, but they have apparently already executed.
Needless to say, most see it as an annoyance not a benefit, me included.
It's not like it's useless but... people tend to hate reviewing LLM output, especially on something like docs that requires proper review (nope, an article and a product are different, an order and a delivery note are as well, and those are the most obvious..).
Code can be subpar or even gross but to the job, but docs cannot be subpar as they compound confusion.
I've even built a glossary to make sure the correct terms are used and kinda forced, but LLMs getting 95% right are less useful than getting 0, as the 5% tends to be more difficult to spot and tends to compound inaccuracies over time.
It's difficult, it really is, there's everything involved from behaviour to processes to human psychology to LLM instructing and tuning, those are difficult problems to solve unless your teams have budgets that allow you hiring a functional analyst that could double as a technical and business writer, and these figures are both rare and hard to sell to management. And then an LLM is hardly needed.
Sometimes they are things I already know but was choosing to ignore for whatever reason. Sometimes it's like "I can see why you think this would be an issue, but actually it's not". But sometimes it's correct and I fix the issue.
I just looked through a couple of PRs to find a concrete example. I found a PR review comment from Codex pointing out a genuine big where I was not handling a particular code path. I happened to know that no production data would trigger that code path as we had migrated away from it. It acted as a prompt to remove some dead code.
Looks bad: https://forum.cursor.com/t/font-on-the-website-looks-weird/1...
Is corporate English becoming a form of newspeak and will significantly diverge from regular English over the next 100 years?
I was pretty worried about Cursor's business until they launched their Composer 1 model, which is fine-tuned to work amazingly well in their IDE. It's significantly faster than using any other model, and it's clearly fine-tuned for the type of work people use Cursor for. They are also clearly charging a premium for it and making a healthy margin on it, but for how fast + good it's totally worth it.
Composer 1 + now eventually creating an AI native version of GitHub with Graphite, that's a serious business, with a much clearer picture to me how Cursor gets to serious profitability vs the AI labs.
I'm very pro IDE. I've built up an entire collection of VSCode extensions and workflows for programming, building, customizing build & debugging embedded systems within VSCode. But I still prefer CLI based AI (when talking about an agent to the IDE version).
> Composer 1
My bet is their model doesn't realistically compare to any of the frontier models. And even if it did, it would become outdated very quickly.
It seems somewhat clear (at least to me) that economics of scale heavily favor AI model development. Spend billions making massive models that are unusable due to cost and speed and distill their knowledge + fine tune them for stuff like tools. Generalists are better than specialists. You make one big model and produce 5 models that are SOTA in 5 different domains. Cursor can't do that realistically.
I have absolutely no horse in this race, but I turned from a 100% Cursor user at the beginning of the year, to one that basically uses agents for 90% of my work, and VS Code for the rest of it. The value proposition that Cursor gave me was not able to compete with what the basic Max subscription on anthropic gave me, and VS Code is still a superior experience to Claude in the IDE space.
I think though that Cursor has all the potential to beat Microsoft at the IDE game if they focus on it. But I would say it's by no way a given that this is the default outcome.
Composer is extremely dumb compared to sonnet, let alone opus. I see no reason to use it. Yes, it's cheaper, but your time is not free.
This is a pretty dumb statistic in a vacuum. It was clearly 100% a few years ago before CLI-based development was even possible. The trend is very significant.
What are we talking about? Autocomplete or GPT/Claude contender or...? What makes it so great?
Cursor has been both nice and awful. When it works, it has been good. However for a long time it would freeze on re-focus and recently an update broke my profile entirely on one machine so it wouldn't even launch anymore.
Kilocode with options of free models has been very nice so far.
Now, would I prefer to use vs code with an extension instead? Yes, in the perfect world. But Cursor makes a better, more cohesive overall product through their vertical integration, and I just did the jump (it's easy to migrate) and can't go back.
I also like how cursor is model-agnostic. I prefer codex for first drafts (it's more precise and produces less code), for Claude when less precision or planning is required, and other, faster models when possible.
Also, one of cursor's best features is rollback. I know people have some funky ways to do it in CC with git work trees etc, but it's built into cursor.
I’ve tried picking up VSCode several times over the last 6-7 years but it never sticks for me, probably just preference for the tools I’m already used to.
Xcode’s AI integration has not gone well so far. I like being able to choose the best tool for that, rather than a lower common denominator IDE+LLM combination.
For backend/application code, I find it's instead about focusing on the planning experience, managing multiple agents, and reviewing generated artifacts+PRs. File browsers, source viewers, REPLs, etc don't matter here (verbose, too zoomed-in, not reflecting agent activity, etc), or at best, I'll look at occasionally while the agents do their thing.
I use VS Code, open a terminal with VS Code, run `claude` and keep the git diff UI open on the left sidebar, terminal at the bottom.
Now that there's MCP, the community will out-innovate anything a single company can do in terms of bolting on features. It's easy enough to get all the LSP integration and stuff into Claude code.
So it all comes down to model differentiation. Can cursor compete as a foundation model creator? Maybe, but even so, that's going to be a very tough market. Margins will be razor thin at best. It's a commodity.
Anyway, the last thing I would want if I were them is to keep worrying about maintaining this IDE themselves.
But also, 90% of the time if I'm using an IDE like VSCode, I spend most of my time trying to configure it to behave as much like vim as possible, and so a successful IDE needn't be anything other than vim to me, which already exists on the terminal
A simple text interface, access to endless tools readily available with an (usually) intuitive syntax, man pages, ...
As a dev in front of it super easy to understand what it's trying to do, and as simple as it gets.
Never felt the same in Cursor, it's a lot of new abstractions that dont feel remotely as compounding
(I work at Windsurf but not really intended to be an ad I’m just yapping)
Even Emacs nuts like me can use agents natively from our beloved editor ;) https://xenodium.com/agent-shell-0-25-updates
I can't randomly throw credits into a pit and say "oh 2000$ spent this month whatever". For larger businesses I suspect it is even worse.
If they had a 200$ subscription with proper unlimited usage (within some limits obviously) I would have jumped up and down though.
Relatively heavy cursor usage in my experience is around 100USD/month. You can set a limit to on demand billing.
I don't understand the "within some limits" people ask for.
If we use a service to provide value, and it is worth the value it provides, why would we ever accept a limit or cap? We want to stop adding value until next calendar month?
Or if the idea is $200 plus overages, might as well just be usage based.
Imagine a rental car that shut off after 100 km instead of just billing 20 km overage to go 120 km. Would you be thrilled for a day of errands knowing the hard cut off? Or would you want flex? You go 60 km out, 40 km back; now it's not worth paying to drive the last 20? If that's the case, probably should have walked the whole way?
Perhaps not a terrible analogy if some devs think of using these models like hitchhiking. Mostly out for the hike but if I can get an Uber now and then for $200/month, then I can do some errands faster, but still hike most places…
OR, hitchhikers don't think they need that much, they only run an errand a week, in which case, back to usage pricing, don't pay for what you don't use.
- - -
As an example: The primary limiter for our firm's wholesale adoption of Anthropic are their monthly caps. The business accounts have a cap! WTH, Anthropic, firms shouldn't LLM review code for the last week or two of the month? It can't be relied on.
To be clear, there's no cap on the usage per se, the cap is at the billing, even if you have it on a corp card that recharges fully constantly, it can tick over at $1500/day for 3 days, then halfway through day 4, it won't recharge again, because you hit $5k/month limit.
If you write to them and ask (like the error messages tells you) they say: Move to Enterprise, it's X users. Well, no, we don't have X people? Sure, but Enterprise is X users. What if we buy empty seats? Um...
(The simplest explanation is that $5k/month really burns more than $5k/month of costs so every API call loses them money, and they'd rather shepherd people to occasional subscription usage where they train them to leave it idle most of the time. Fine, offer usage at cost instead of loss, see who bites.)
Meanwhile, we use unlimited from their competitors, and have added several other ways to buy Anthropic indirectly, which seems weird they'd want to earn less per API call but someone somewhere is meeting their incentives I guess.
Also Cursor's dataset of actual user actions in coding and review is pure gold.
Fascinating.
As a person who *loathes VS Code* and prefers terminal text editors, I find Cursor great!
Maybe because that I have zero desire to customize/leverage Cursor/VS Code.
Neat. Cursor can do what it wants with it, and I can just lean into that...
Are you sure entire opinion is not just centred around that fact? Sounds like it
The UX of IDE integration with the existing VSCode plugins and file manager… it’s not even close to the same. Some people just get comfortable with what they are comfortable with
I’ve tried every popular agent IDE, but none of them beat Cursor’s UX. Their team thought through many tiny UX details, making the whole experience smooth like a butter. I think it’s a huge market differentiator.
Also their own composer model is not bad at all.
I don't envy startups in the space, there's no moat be it cursor or lovable or even larger corps adopting ai. What's the point of Adobe when creating illustrations or editing pics will be embedded (kinda is already) in the behemoth's chat it's?
And please don't tell me that hundreds of founder became millionaires or have great exits or acquihires expecting them. I'm talking about "build something cool that will last".
A VSCode fork with AI, like 10 other competitors doing the same, including Microsoft and Copilot, MCPs, Vs code limitations, IDEs catching up. What do these AI VsCode forks have going for them? Why would I use one?
More specific models with faster tools is the better shovel. We are not there yet.
Graphite is a really complicated suite of software with many moving pieces and a couple more levels of abstraction than your typical B2B SaaS.
It would be incredibly challenging for any group of people to build a peer-level Graphite replacement any faster than it took Graphite to build Graphite, no matter what AI assistance you have.
Much respect to what have you have achieved in a short time with graphite.
A lot of B2B SaaS is about tones of integrations to poorly designed and documented enterprise apps or security theatre, compliance, fine grained permissions, a11y, i18n, air gapped deployments or useless features to keep largest customers happy and so on and on.
Graphite (as yet) does not any of these problems - GitHub, Slack and Linear are easy as integrations go, and there is limited features for enterprises in graphite.
Enterprise SaaS is hard to do just for different type of complexity
What I do not understand is that if a high level staff with capacity can produce an 80% replacement why not assign the required staff to complete that last 10% to bring it to production readiness? That last 10% is unnecessary features and excess outside of the requirements.
I hate the unrealistic AI claims about 100X output as much as anyone, but to be fair Cursor hasn't been pushing these claims. It's mostly me-too players and LinkedIn superstars pushing the crazy claims because they know triggering people is an easy ticket to more engagement.
The claims I've seen out of the Cursor team have been more subtle and backed by actual research, like their analysis of PR count and acceptance rate: https://cursor.com/blog/productivity
So I don't think Cursor would have ever claimed they could duplicate a SaaS company like Graphite with their tools. I can think of a few other companies who would make that claim while their CEO was on their latest podcast tour, though.
Also, graphite isn't just "screenshots"; it's a pretty complicated product.
My usually prefer Gemini but sometimes other tools catch bugs Gemini doesn't.
As someone who has never heard of Graphite, can anyone share their experience comparing it to any of the tools above?
> "Will the plugin remain up? Yes!"
sweet summer child.
The results of a diff-only review won't be very good. The good AI reviewers have ways to index your codebase and use tool searches to add more relevant context to the review prompt. Like some of them have definitely flagged legit bugs in review that were not apparent from the diff alone. And that makes a lot of sense because the best human reviewers tend to have a lot of knowledge about the codebase, like "you should use X helper function in Y file that already solves this".
Then it can run `git diff` to get the diff, like you mentioned, but also query surrounding context, build stuff, run random stuff like `bazel query` to identify dependency chains, etc.
They've put a ton of work into tuning it and it shows, the signal-to-noise ratio is excellent. I can't think of a single time it's left a comment on a PR that wasn't a legitimate issue.
You might want to look at existing products in this space (Cursor's Bugbot, Graphite's Reviewer FKA Diamond, Greptile, Coderabbit etc.). If you sign up for graphite and link a test github repo, you can see what the flow feels like for yourself.
There are many 1000s of engineers who already have an AI reviewer in their workflow. It comments as a bot in the same way dependabot would. I can't share practical lessons, but I can share that I find it to be practically pretty useful in my day-to-day experience.
Yes, you definitely need the project's context to have valuable generations. Different teams here have different context and model steering, according to their needs. For example, specific aspects of the company's architecture is supplied in the context. While much of the rest (architecture, codebases, internal docs, quarterly goals) are available as RAG.
It can become noisy and create more needless review work. Also, only experts in their field find value in the generations. If a junior relies on it blindly, the result is subpar and doesn't work.
If that's not the concern, then what's the big deal?
Good thing that Copilot is not the dominant tool people use these days, which proves that (in some cases) if your product is good enough, you can still win an unfair competition with Microsoft.
If you've used Graphite as a customer for any reasonable period of time or as part of a bigger enterprise/org and still think our app's particular integration with GH is easy... I think that's more a testament to the work we've done to hide how hard it is :)
Most of the "hard" problems we're solving (which I'm referencing in my original comment) are not visually present in the CLI or web application. It's actually subtle failure-states or unavailability that you would only see if I'm doing my job poorly.
I'm not talking about just our CLI tool or stacking, to clarify. I'm talking about our whole suite, especially the review page and merge queue.
What kind of enterprise SaaS features do you wish you had in Graphite? (We have multiple orgs with 100s-1,000s of engineers using us today!)
This is something GitHub should be investing time in, it’s so frustrating.
I've been going through a lot of different git stacking tools recently and am currently quite liking git-branchless[1] with GitHub and mergify[2] for the merge queue, but it all definitely feels quite rough around the edges without first-party support. Especially when it comes to collaboration.
Jujutsu has also always just seemed a bit daunting to me, but this might be the push I needed to finally give both jj and tangled a proper try and likely move stuff over.
I've been using composer-1 in Cursor for a few weeks and also switching back and forth between it, Gemini Flash 3, Claude Opus 4.5, Claude Sonnet 4.5 and GPT 5.2.
And you're right it's not comparable. It's about the same quality of code output of the aforementioned models but about 4x as fast. Which enables a qualitatively different workflow for me where instead of me spending a bunch of time waiting on the model, the model is waiting on me to catch up with its outputs. After using composer-1, it feels painful to switch back to other models.
I work in a larg(ish) enterprise codebase. I spend a lot of time asking it questions about the codebase and then making small incremental changes. So it works very well for my particular workflow.
Other people use CLI and remote agents and that sort of thing and that's not really my workflow so other models might work better for other people.
The Copilot version of this is just fucking terrible at suggesting anything remotely useful about our codebase.
I've had reasonable success just sticking single giant functions into context and asking Sonnet 4.5 targeted questions (is anything in this function modifying X, does this function appear to be doing Y) as a shortcut for reading through the whole thing or scattershot text search.
When I try to give it a whole file I actually hit single-query token limits.
But that's very "opt-in" on my part, and different from how I understand Cursor to work.
And when I open it in the parent directory of a bunch of repos in our codebase, it can very quickly trace data flow through a bunch of different services. It will tell me all the files the data goes through.
It's context window is "only" 200k tokens. When it gets near 200k, it compresses the conversation and starts a new conversation..... which mostly works but sometimes it has a bit of amnesia if you have a really long running conversation on something.
Imaginary situation: People are using claude instead of cursor, and you can run claude in a terminal, so this is going back to the days of not using an IDE for the people that do it.
Straw man shake down: Terminal based development like vim and emacs are old and shit, and we moved away from that for a reason, and so (although totally unrelated) this means 'using claude' means going back to using a terminal for everything, which is similarly old and shit.
...but, obviously wrong.
- There's a claude desktop app that isn't done via the terminal.
- Agents use the terminal/powershell to do lots of things, even in cursor because that's the only way to automate some things, eg. running tests.
- Terminal environments like vim and emacs are ides. :face-palm:
- It literally makes no difference what interface you copy and paste your text prompt into and then walk off to get a coffee in agent mode.
Anyone who's seriously arguing that IDE integrated LLM chat windows somehow beat command line LLM chat windows is either a) religiously opposed to the terminal window, or b) hasn't actually tried using the tools.
...because, you'll find it makes no difference at all.
Why is cursor getting involved with graphite? ...because the one place where is makes a difference is reviewing code, where most CLI based tools (eg. `git diff`) are just generally inferior to visual integrated code review tools.
You know what that is?
An acknowledgement that cursor, in terms of code generation has nothing that qualifies as the 'special sauce' to use it over any other tool. CLI or not.
So they're investing in another company that actually has a good, meaningful product.
I am betting it won't.
By the way, there are OS APIs, I am yet to write a CLI driven agent, as part of iPaaS deployments, which are basically SaaS IDEs.
vi and Emacs are certainly not IDEs, they are programmer editors, although with enough effort they may pretend to be one.
I don't even like using CLI, in fact I hate it, but I don't use CLI - Claude does it for me. Using for everything: Obsidian vault, working on Home Assistant, editing GSheets, and so much more.
Right now VSCode can do things that Cursor cannot, but mostly because of the market place. If Cursor invests money into the actual IDE part of the product I can see them eclipsing Microsoft at the game. They definitely have the momentum. But at least some of the folks I follow on Twitter that were die-hard Cursor users have moved back to VSCode for a variety of reasons over the last few months, so not sure.
Microsoft itself though is currently kinda mismanaging the entire product range between GitHub, VS Code and copilot, so I would not be surprised if Cursor manages to capitalize on this.
Which is what I was mentioning elsewhere. They build huge models with infinite money and distill them for certain tasks. Cursor doesn't have the funding, nor would it be wise, to try to replicate that.
That being said, surely the point here is about "agent driven development" vs "ai autocomplete". As they say, whether you type your command into a web window or a terminal window presumably doesn't change the flow that much.
everyone is staying on to keep making the graphite product great. we're all excited to have these resources behind us!
It's happened so many times that it's just part of how we do business, unfortunately.
If Cursor wants to re-allocate resources or merge Graphite into to editor or stagnate development and use it as a marketing/lead gen channel, it will for the business.
Anything said at time of acquisition isn’t trustworthy. Not because people are lying at the time (I don’t think you are!) but because these deals give up leverage and control explicitly. If they only wanted tighter integration, they could fund that via equity investment or staffing engineers (+/- paying Graphite to do the same.) Companies acquire for a reason and it isn’t to let the team + product stay independent
We know this isn't what all of you want to hear, and we've spent the last year really evaluating this deeply. At the same time, we're glad you're part of our journey to the future of agentic AI and we think you'll find it's the best alignment and fit for you, too, long-term.
Something like 90th percentile usage is what I'd call relatively heavy.
Bake that into the workflow some other way.
How does that work? Multiple agents grepping simultaneously?
LLMs are inherently single-threaded in how they ingest and produce info. So, as far as I can gather from the description, either it spawns sub-agents, or it has a tool dedicated for the job.