Rosebud AI

Browser-side AI game creation platform, using natural language to generate 2D/3D web games that can be played and shared in one sentence

Vibe Coding game generation

https://rosebud.ai/

Visit Website

In-depth Report

Rosebud AI is a browser-side AI game creation platform for zero-based users, located at rosebud.ai. Users describe their ideas in natural language, and the AI assistant "Rosie" turns the ideas into 2D/3D web games that can be tried immediately and can share links. It compresses writing code, generating materials, and deploying online into just a few minutes. It is positioned as "the prototype of Roblox in the game development industry" and a representative product of vibe coding.
The company was founded by Lisha Li, a Ph.D. in statistics from Berkeley. It came out of Y Combinator and was named to the Forbes AI 50. It started with the game material generation tool PixelVibe in the early days. It launched the core AI Game Maker in 2024 and released the exclusive basic model CRAI for games in the same year. Lisha Li will step down as CEO in 2026 and join the well-known venture capital a16z as an investment partner, focusing on AI and infrastructure. According to third-party estimates, its annual recurring revenue is about US$2.7 million, and its cumulative financing is about US$8.8 million to US$15 million. Investors include Foundation Capital, Khosla Ventures, Foothill Ventures, Animoca Brands and YC. Angel rounds include Jeff Dean, Ilya Sutskever, Andrej Karpathy and others. Need to be reminded that there is another "Rosebud" diary app on the market with the same name (rosebud.app, founder Chrys Bader, received a $6 million seed round in 2025). The two are completely different companies and products, and it is easy to confuse them during evaluation and retrieval. This report only talks about the game creation platform.
The core capability is to directly generate playable games using text or speech. The user writes something like "a dragon that chases the player" in the input box, and Rosie generates JavaScript game code, built-in materials, and game loops. It takes a few minutes to take out the prototype and click Apply to update it in real time. The platform has built-in PixelVibe (generating characters and textures), AI Storybook (generating plots), TokkingHeads (making static portraits speak), and Manyworlds (3D skyboxes and scenes). It also supports voice commands, multiplayer templates, community Remix, and one-click cloud deployment and sharing. The technical base is browser-side JavaScript. Different materials mention Phaser and Three.js respectively. The overall technology is mainly 2D and 3D is the primary capability. Its biggest advantage is its "fastness" - almost no one can match it in prototypes, classroom demonstrations, and game jam scenes. Its shortcomings are also the price of "fastness": complex logic can easily get out of control, you may need to be prompted repeatedly to change the size of an elf, and the 3D physics feel is wavy.
With Free Plus Pro subscription. The latest public price (May 2026) is Free US$0, Pro Creator US$19.99/month, Enterprise customized pricing; heavy calculation tasks (3D rendering, complex animation) are additionally charged with Spark Credits computing power points. The free file has limited generation times, no commercial copyright, and no open code; the Pro file has commercial copyright and code visibility. In the early days, I used the Liftoff / Orbit / Galaxy points tier (approximately $12.99 / $39.99). The pricing has gone through several rounds of adjustments, and the trend is towards the dual track of "subscription + computing power points". Discounts available for students and users with disabilities.
The community as a whole was amazed by the "playable prototype in a few minutes". The matching of material styles and the friendliness of the Discord community were repeatedly praised. The negatives are concentrated in the debugging process: complex bugs often require reading and correcting the code yourself, which violates the "zero code" promise; fine animation control is insufficient; AI-generated art occasionally has repeated patterns. Some users on Reddit complained that "it took several hours to change the size of the sprite animation, and I almost lost my hair."; a GitHub discussion believed that "the abstraction layer is too limited for extremely complex AAA, but it is unparalleled in quick-to-market prototypes."
The media generally regards Rosebud as the most mature member of the prompt-to-game circuit. It is good at small arcades, top-down adventures and simple platform jumps, and the exported code can be taken away and continued to be modified. Among competitors on the same track, Nilo (Summer Engine) focuses on real 3D and multi-person collaboration, and can be exported to Roblox/Unity; Astrocade focuses on social remixing and emphasizes "60 seconds of fun"; Websim generates entire interactive web pages but is the least reliable; GDevelop, Godot, and Construct 3 are "graduate" engines that can truly produce products. Rosebud's boundaries are exactly those of web game frameworks: 2D, browser-scale, weightless systems.
One is the ceiling of about 2,500 lines of code - when complex projects hit the upper limit of the large model context window, AI begins to lose early code and introduce new bugs that cannot be fixed. This is the limit of capabilities, not bugs. The second is pure web deployment, with no native desktop or mobile export. If you want to get on Steam or make real 3D products, you have to find another way. Third, free files do not have commercial copyrights, and they do not have the right to monetize their works if they become popular. This is often told too late. Fourth, Spark Credits will quietly add costs under 3D heavy load. The fifth is a solo tool without team collaboration (except Enterprise). In addition, billing and refund disputes with the diary app of the same name are often mistakenly attributed to this brand, and need to be screened when searching.
It is suitable for enthusiasts with no basic knowledge, game design students, independent developers who want to quickly verify gameplay, teachers for classroom demonstrations, and Game Jam players who want to make a playable thing tonight. Not suitable for those aiming for a Steam-ready, true 3D, salable finished product, nor for teams requiring robust 3D physics, team collaboration, or commercial-grade polish. Alternatives: If you want real product, choose GDevelop / Godot, if you want strong 3D and multiplayer, choose Nilo, if you want a spark of inspiration, choose Websim / Astrocade.
Rosebud AI has done a job of "making a game just by saying a word" better than expected. Once the link is sent to the other party, it can be played in three seconds. It is currently the most cost-effective entrance in prototyping and teaching scenarios. However, it is essentially a prototyping tool rather than an engine. The 2,500-line ceiling, pure webpage deployment, and commercial copyright threshold determine that it is good at igniting ideas, rather than carrying ideas into products.

User Reviews

JRiveraII_208

—

Absolutely.

CherylClarkK52

—

There is no going back.

tfmvj4q8_io

—

There are only so many free slots per week.

James.Hicks_2022

—

It takes ten minutes to create a playable thing, which is outrageous.

Jerry_Nguyen8

—

The voxel application uses the Noa engine and Babylon.js, which is quite smooth for Minecraft.

Jack.BakerZ

—

Pure web deployment, no desktop or mobile version, so if you want to get on Steam, you have to find another way.

CAlop

—

Pro costs $19.99 a month, which is pretty student-friendly.

Clara980

—

The free version can only play like that, if you want to be more serious, you still have to open the Pro.

Cheryl_RuizIII

—

There is a real case in our circle. Shawn used pure vibe coding to create the first game Flow State with a Steam page, and he also won the Game Jam runner-up and a $2,000 bonus. This shows that this thing is not just a toy, but can actually be sold in the real market for money. The premise is that you can accept its code limit and are willing to complete it yourself based on the prototype, rather than expecting to produce a finished product in just one sentence.

Kevin.Ramos_X6

—

The official website says that it has 1.5 million users. This growth is indeed rapid. It seems that there is a real demand for "making games with just one sentence".

Willie838

—

The Showcase in the community is too popular. There are a lot of games that can be remixed directly. Novices can learn faster by following the remix than by reading the documentation.

AdemSandalcı

—

It is for students to use in classes. Children with zero basic knowledge can also make small games in the browser, which is much more intuitive than Unity.

JesseTorres_20242

—

I made a horror shooting game, and the AI was able to fix bugs based on screenshots and descriptions, and I was able to create a complete demo in one day. It was quite magical.

CherylNelson_Plus

—

Game Jam artifact, I need it to hand over something playable tonight. Once the link is sent to the opponent, it can be played in three seconds.

Grace_GonzalesX

—

You still have to debug complex bugs yourself. The so-called zero code does not work in 20 to 30% of cases.

LambdaLPCox

—

Changing the elf size was stuck for me all afternoon, and I kept getting prompts and prompts, but nothing seemed right.

MHicksII70

—

Material generation and code generation are in the same interface, so there is no need to fiddle around with external tools, which is very friendly to lazy people.

ISbut

—

I used it to make a 3D mini-game in which a UFO catches a cow. The first ten minutes were really fun. The UFO would move, the tractor beam would suck up the cow, and the sucked cow would disappear. But once you want to add scores, timing, and multiple levels, things get messed up: the modified code often doesn't take effect, regeneration washes out the previous logic, and the 3D physics doesn't work well. In the end, it can only be used as a prototype to demonstrate to friends, and it is still far from the stage of real playability.

AbigailButler_Max

—

Rosie writes the basic structure quickly, but the more complicated state machines start to mess up, and in the end she has to read and modify the code herself.

DSandersX61

—

The bottom layer of the multiplayer template is Supabase. It is very fast to add an online layer, but it is only to the level of amateur online games. Don’t expect large-scale online games.

DrEleonoraŠotra_x

—

Frankly speaking, the first fifteen minutes were like magic, but the next three hours were reality. It's best suited for classroom demos, game jams, and one-off prototypes where you want to iterate on a finished product, and the ceiling will come faster than you think. My suggestion is to treat it as a learning tool, which is the most cost-effective; if you really want to use it as an engine to carry a complete product, plan your exit path as early as possible.

Ronald_JamesIII

—

Let’s talk about a few pitfalls that are easy for novices to step into: First, there is a code ceiling of about 2,500 lines. When complex projects hit it, the AI will start to lose previous logic and introduce new bugs that cannot be fixed; second, the free file has no commercial copyright at all, and there is no right to monetize the work if it becomes popular; third, Spark Credits will quietly pile up money under heavy 3D load, and at the end of the month it will be higher than the subscription fee. Think carefully before deciding whether to join.

NMurphy_776

—

In terms of speed, Rosebud really has no rivals. It is unmatched in making a prototype that can be quickly put to market. Veterans also recognize this. But what you want to do is a high-fidelity AAA masterpiece, so this layer of abstraction will get in the way. Experienced people will be more comfortable using C++ or C#. Its positioning is actually very clear: it is for people who don’t want to touch grammar at all and just want to turn ideas into playable things. Don’t compare its depth with Unity.

Rosebud AI

In-depth Report

User Reviews

Similar Products