Rosebud (YC S19) – Turn game descriptions into browser games

Hi everyone! I'm Lisha, the founder of Rosebud AI ( We're building a platform to help users go from description to code to game. We aim to make game creation accessible to non-technical creators, so our UI provides explanations alongside the generated code.

Users have created a diverse range of games on Rosebud, including top-down RPGs, AI companions, and 3D obstacle courses, all within a few hours and sometimes minutes. Here are some examples you can play and clone (to start your own project).

* Anime Jester Companion: * Chat and Care for your Digital Puppy: * Sphere Sync (3D game: align the sphere with the right color): * Basketball: * Neon Waltz Generative Art: * Chat with Deku from My Hero Academia:

A simple way to think about Rosebud is ChatGPT + Midjourney + Replit. ChatGPT, because we give users a chat interface for this code editor so they can describe the game they want to make and generate game code; Midjourney, because we let users generate assets inside Rosebud, 2D and 3D, to be used in their games; And Replit, because Rosebud includes a browser based code editor that lets you deploy your game instantly.

Sometimes, users generate a code base from scratch via prompts. Often a simpler place to start is to modify (“clone”) an existing project on Rosebud. In both cases, we need to eventually convert user descriptions and modifications of the game into edits and changes to the codebase. To solve this problem, we had to experiment heavily with using LLM agents in production. Our agent framework tries to follow the instructions of user prompts by deciding when and whether to call upon a number of generative models (some for code generation, some for asset generation, some for character dialogue, and some for game ideas). It also must decide where to insert code snippets when it generates them. Often, a user is asking for ideas or something too vague, and our agent has to decide when to ask for feedback and clarifications.

Not surprisingly, if we impose more constraints, on both the programming framework and game genres supported, our agent will perform better. However, the constraints on the types of games users can make and frameworks we want to support also constrains how flexible our platform is. Balancing these two factors, we decided to only support browser-based, JavaScript frameworks and focus on supporting AI NPCs that use LLMs themselves for dialogue and actions. This allows us to create abstractions that enable the agent to alter the codebase more successfully and guide the creator towards a more successful experience. Furthermore, we found that our beta testers are very creative with making AI character based games, and the resulting game is usually fun for players.

How does Rosebud differ from Roblox, Unreal, or Unity simply adding a co-pilot? Incumbent game engines optimized their user-flow and tech stack before the advent of generative AI, and many of their user-flows are well established. We have the advantage of designing this game creation flow from the ground up. It's not just about adding code completion to an existing code editing app and including asset plugins. Such an approach wouldn't fully harness the power of LLMs. We have a chat-first interface, and having identified the limitations of agents, we can create more safeguards for users where failure is likely. Our approach will make it possible for non technical creators to also contribute to making games. Check it out for yourself!

To try Rosebud: (1) head over to for access to our Discord beta tester channel and a special role. (2) then go to and use the code HelloHN to get immediate access. We have an array of trending projects that users can clone and mod to get started, including various character chat based games.

Here’s a video onboarding of Rosebud in action:

Re business model, we plan on following in Roblox’s footsteps, i.e. keep it free for developers and take a cut of what they can charge users. Since AI tools cost more from usage than just hosting, we may have to evolve that model and see what the unit economics are (and separate a premium versus free tier for devs).

(Oh and in case you’re wondering why a YC S19 startup is launching now: we basically pivoted. We were always in consumer generative AI, but focused on images until this spring, but always wanted to focus on games–Rosebud is in fact a reference to the cheat code in The Sims. When code gen got good enough this year to work for UGC in gaming, we decided the time had finally come and switched.)

Some encouraging user feedback from our beta: “I have done some modding before, and I must say, this is much easier. Even when I occasionally need to code, the AI can answer all my questions and tell me how to achieve what I want. Normally, I would have to conduct numerous Google searches. What you guys have created is truly amazing.” “I’ve used Chat GPT to help me code simple games in Unity. This seems more connected and easier to work through.” “This is fascinating. This is ** amazing. Yeah, I know it's obviously early on, but already works for rad generative art. I'll say that much.” “Can finally call myself a game developer lol. Damn that sounds so good.”

We’re a small team working on this for the last few months, so a lot of things are far from perfect. Constructive feedback is very welcome!

Get Top 5 Posts of the Week

best of all time best of today best of yesterday best of this week best of this month best of last month best of this year best of 2023 best of 2022 yc s24 yc w24 yc s23 yc w23 yc s22 yc w22 yc s21 yc w21 yc s20 yc w20 yc s19 yc w19 yc s18 yc w18 yc all-time 3d algorithms animation android [ai] artificial-intelligence api augmented-reality big data bitcoin blockchain book bootstrap bot css c chart chess chrome extension cli command line compiler crypto covid-19 cryptography data deep learning elexir ether excel framework game git go html ios iphone java js javascript jobs kubernetes learn linux lisp mac machine-learning most successful neural net nft node optimisation parser performance privacy python raspberry pi react retro review my ruby rust saas scraper security sql tensor flow terminal travel virtual reality visualisation vue windows web3 young talents

andrey azimov by Andrey Azimov