Tusk is an AI agent that helps product teams complete UI changes from ticket to pull request. Automate away grunt work like minor bug fixes and copy changes to increase customer NPS without bothering your software engineers.
Hey everybody!
@sohil_kshirsagar, @jytan and I are proud to launch Tusk, an AI agent that helps PMs and product engineers make UI improvements without needing to write code. 🐘 🎉
We launched our beta earlier this year with a select group of companies while we were going through the YC W24 batch. We’ve since battle-tested our AI across a variety of production tasks in mature, complex codebases.
It’s been incredibly exciting working on the frontier of what LLMs can do in the wild. We’ve put a lot of R&D cycles into getting LLMs to make the right changes in the right files. Our state-of-the-art agent refers to an abstract semantic graph of your codebase and learns from your past PRs and code reviews to generate high quality code.
💨 Tusk integrates with ticketing software like Jira, Linear, Notion, and GitHub so it takes one click to create a pull request for your UI tickets
🐛 Tusk comes with out-of-the-box Figma, Loom, Jam, and Bird Eats Bug integrations that pull context from external apps to generate high-quality code
🧑🔬 Our agent runs your CI checks on its code output and automatically iterates on a branch until it passes all your checks
🧠 Our agent addresses and remembers feedback from any code reviews left by human software engineers
We’re a firm believer that you shouldn’t need to bother your software engineers to fix padding on a modal or change the text of a header. Evidently we’re not alone. Tusk is now helping high-growth companies, backed by the likes of Y Combinator, Andreessen Horowitz, General Catalyst and Tiger Global, save $36K in engineering hours a year.
Try Tusk for free on your own repo today: https://usetusk.ai/
Feel free to ask your questions here or reach out to me at marcel@usetusk.ai. :)
@sohil_kshirsagar@jytan@dash4u thanks for the kind words! Yes, we do support more complex UI changes like adding a net new component or changing the interaction of an element. The popular use case is fixing bugs since that has more of an impact on customer NPS.
If you go to https://usetusk.ai/ and click on "Customers" in the website header, you can see real examples of customer tickets that Tusk has created a merged PR for.
@sohil_kshirsagar@jytan@kjosephabraham thanks for the support Joseph! We had the idea while working as a PM and product engineer at our previous jobs.
As a PM, I'd want to get UI polish tickets resolved before a launch or fix a minor bug to make a customer happy. But as Sohil and Jun Yu can attest, these types of tasks often take time away from already over-burdened engineers. We'd been experimenting with LLMs since GPT-3 (I was a PM at an AI company) and we saw a clear use case for having an agent automate away these engineering chores.
@sohil_kshirsagar@jytan@marceltan great work folks! When do you think Tusk won't be a good fit? Large code bases? Code that's not well segmented? Large monolithic apps?
Congrats on the launch! This tool seems super helpful for PMs and engineers. How do you see product designers collaborating with PMs and engineers using this tool?
Again, really excited to try this out!
@grant_kalasky thanks for the support Grant! Looking forward to having you give Tusk a spin. :)
Product designers are actually a good fit for Tusk given that design language translates quite directly into frontend dev work. The designer using Tusk would essentially be in the role of a PM assigning a ticket to Tusk and then passing the PR to Engineering for review. Works best in horizontal teams where designers (and not just PMs) have autonomy over getting UI/UX tweaks resolved.
@grant_kalasky Thanks Grant! We've actually noticed that product designers have a lot of UX quality tickets in the backlog (or in their head) they want to get fixed, but they don't get prioritized. Tusk works really well on these tickets, and the fact that designers tend to write great tickets doesn't hurt!
Awesome product with a great team 🚀
I have already used Tusk to make some improvements to our Formbricks app. It was so great to see a high quality PR created in just a few minutes! 🔥
I use cursor everyday for stuff like this. but It needs a lot of handholding for smaller tweaks, as the edge specific cases are not usually something general knowledge LLMs are gd at solving. But the LLM will often give some nice ideas that can be tweaked and made to work. I guess Tusk could be useful in the scenario as well. Maybe not 100% solves it, but gd enough for an engineer to take it further? Whats the experience here. I know the marketing will want to say its a one stop solution that solves it all 100%. but that's just not feasible with todays level of LLM's. Whats would the honest marketing say? 😬
@sentry_co hey André, thanks for the question! Our unassisted PR merged rate for smaller tickets is 71% for codebases of a good fit. You're right that there are tasks where Tusk creates a draft PR that's 80% of the way there. An engineer can then checkout the branch and complete the rest.
In-IDE tools are great for writing code faster; we use them ourselves. Tusk is different in that we intentionally live outside of the IDE. We believe that humans shouldn't have to be in the loop when using LLMs for frontend grunt work. Our customers have Tusk running on their chore tickets in the background while they work on other tasks.
@marceltan Looking forward to follow this. It will get very interesting when the LLM's improve and more PRs can be made automatically and completed as is. 💪
@sentry_co appreciate it André! We post product updates regularly on our LinkedIn (both our personal profiles and the Tusk company page). Let me know if we can be helpful to you as you're building Sentry. All the best for your upcoming launch :)
@marceltan Hey it's amazing to see Tusk launched on Product Hunt! Congratulations, Tusk is a fantastic product and it's great to see it getting the recognition it deserves
@copaceticd thanks so much Dan! It's been a joy for @sohil_kshirsagar and I to be working with you to get PRs merged for Prefix. Excited to share some upcoming features related to automated testing with you very soon!
Tusk has been amazing for us at replo.app! Really cool to outsource busywork to an AI so that our engineers don’t have to waste time on stuff they shouldn’t need to
@noahsark769 appreciate the review Noah! @sohil_kshirsagar and I have really enjoyed working closely with you, Yuxin, and the rest of the Replo team 🤩 we're pumped to show you our upcoming automated testing features. Want to make this the best possible product for you.
A killer product with a killer team, congrats on the launch!
Super curious where you see this heading for languages that fall somewhat out of the major distribution such as Swift?
@vatsal_manot thanks for the kind words, Vatsal! Candidly, optimizing for Swift is not a focus for Tusk right now. With that said, we are looking into incorporating external docs as context for Tusk if an issue requires it, which would boost the agent's ability to generate Swift.
Hi @marceltan
Congrats on the launch! Seems like Tusk is a useful nice tool :)
I created this shortlink for you that will send users directly to this PH page, in case you feel like the PH link is too long :)
Here is the link: https://shor.tf/tusk-ph
@yash3 thanks for the questions! Tusk works well with popular frontend frameworks like React, Angular, and Vue.js. We've typically seen that B2B2C or consumer marketplace companies get the most mileage out of Tusk because these companies get a constant stream of customer-reported bugs and feature requests flowing into their sprints. Their PMs/engineers will offload these chores to Tusk so they can focus on higher-priority work.
Congrats on the launch :) I’m curious what the most complex tasks are you’ve seen Tusk successfully solve? Also curious how far away you think we are from Tusk being able to fully replace a product engineer?
@rohan_varma1 thanks Rohan! While backend isn't the focus right now, we've recently used Tusk internally to write the backend code and LLM prompt to better handle code snippet selection when the AI agent addresses code reviews. We merged that PR with no human commits.
Seems like we're still ~3 years away from coding agents being able to perfectly do the full job of a product engineer. The hard part isn't generating the code, it's actually finding all the relevant context and then reasoning through said context.
congrats on the launch Marcel. This is full of potential, I'll have to try it out with my repo to really testify the effectiveness, but I really think it's the future of AI Agent. Way to be ahead of the game, congrats!
@danielwchen thanks for the kind note Daniel! Please do try it out on your repo and let me know your thoughts. My line is always open at marcel@usetusk.ai.
@ken_jyi_lim thanks for asking! For our most active customers, Tusk is contributing 48.9% of all monthly merged PRs. Caveat here is that we're creating smaller PRs. Across ICP customers, we've seen that 45.5% of Tusk's PRs are merged without any human commits at all. The remainder consist of PRs where an engineer will checkout the branch and work off of it manually.
@rumiz Thank you Rumi! If hallucinations results in build errors or lint issues, we fix them with a GitHub workflow that automatically addresses these feedback. To guard against syntactic hallucinations, we check the code against an abstract syntax tree. For "reasoning" hallucinations, we do our best to avoid them by using an abstract semantic graph and long-term memory (Tusk's merged and closed PRs on your tickets over time).
Congrats on the launch @marceltan@sohil_kshirsagar! Little UI fixes always get pushed to the back burner at my company. What kind of grunt work do you find Tusk is best at?
@harrison_paradise Thanks Harrison! Totally agree, it's especially frustrating when these UI fixes are things customers are complaining about. Tusk works best on tickets that have enough context (we support images, Loom links, Jam links, etc.) and originate from the frontend!
Whoa this is pretty sweet! As a PM, I can see how this can take care some technical debt and small bugs, so engineering team can stay on top of the big rocks. Love how the AI agent actually does QA work and will tell you if an issue is too complex! Knowing the boundary is so important for AI.
I can see that over time, as the AI learns from the CR, and reading commits and notes, it will get smarter and fine tuned for the codebase.
What does the future look like for Tusk? Running Tusk proactively to discover issues and suggest fixes? I can see many use cases extended from taking care of simple tasks today.
Congrats on the launch @marceltan and team!
@marceltan@tonyhanded Thank you Tony! 100% agree, uncertainty estimation is one of the biggest technical challenges when scaffolding reliable AI products. And yes, Tusk learns from its merged and closed PRs on your tickets.
Proactively discovering issues is not on our roadmap at the moment, but Tusk already has an “auto-triaging” feature, suggesting actionable tickets from your Linear/Jira/Notion board for you to assign to it! The idea is to make solving simple UI tickets from your growing backlog as seamlessly as possible.
Next up, we’re making strides to improve testing — writing unit tests alongside your test suite, automated sanity checks with preview environments so that you can be more confident that changes are correct and spend less time reviewing PRs. We’re also advancing Tusk’s codebase-understanding ability so we can better localize faults and solve them 🚀
Tusk (YC W24)