Rosebud AI Teardown — Text-to-Game Indie Pivot After Three Tries ($40K MRR, $8.8M Raised)

Published 2026-05-16. Last verified 2026-05-16.

TL;DR

Rosebud is a strange case to study because it does not fit the shape of the other reports in this batch. The company has been alive for roughly seven years. It has burned through three distinct product theses. The current one — type a sentence, get a playable browser game — works well enough to sustain a small team, a 17K-member Discord, and an estimated $40K MRR. That is not a venture outcome. It is, however, more than 95% of indie hackers ever reach, and the path Rosebud took to get there is unusually instructive.

I spent an afternoon inside the product. I typed a prompt — "a top-down maze where you collect dropped umbrellas while avoiding rain clouds" — clicked Create, and roughly 90 seconds later I was steering a sprite around a procedurally generated grid. The art was rough. The collision detection was glitchy. But it ran, in a browser, and I had written exactly zero lines of code. That is the product. That is the entire pitch.

The interesting story is not the product. It is the founder's refusal to die. Lisha Li started Rosebud in 2019 out of YC S19 with a thesis that turned out to be wrong (synthetic media for fashion marketing). She pivoted into AI-animated portrait avatars under the brand Tokkingheads (got millions of users, no business model). She raised a $6.6M Series A in late 2021 from Animoca Brands at the peak of the NFT cycle, pointed the product at character-as-IP/Web3 use cases, and watched that thesis evaporate when NFTs collapsed. Then, in January 2024, she launched what the company should have been all along — a text-to-game maker that lets non-developers ship browser games.

Three theses, eight years, $8.8M raised across rounds, and the current company is generating something in the range of $40K MRR with a $29/month Pro tier. The most recent and revealing data point: Lisha announced in mid-2025 she was stepping down as CEO to join a16z's infrastructure team as an investing partner. That is the kind of move you make when the company has stabilized into something that runs without you, but is not big enough to be the rest of your career.

Here is what the indie hacker should take from this:

• You can be wrong twice and still ship something real on the third try.
• A famous-in-AI-circles founder generates customers more cheaply than ads can.
• The technical moat in "AI builds X for you" products is almost zero. The moat is brand, community, and a constrained vertical that the big general-purpose tools (Bolt, Lovable, Cursor) are not optimizing for.
• $40K MRR is not a failure mode. It is a wage-replacing outcome that most YC founders never reach.

This report walks through the three pivots, reverse-engineers the current product, and tells you exactly how to clone the playbook in a different vertical without raising $8.8M or spending eight years on it.

In the Founder Own Words

"Rosebud AI Open Game Protocol ( @ogprotocol ) Build your game with AI, no code needed. OGP handles the publishing + token rewards. You just focus on vibe-coding."

• @rosebud_ai, 2025-10-23 (source)

"Call of Duty: Black Ops 7...Could the game look like THIS? Create with Rosebud AI → https:// rosebud.ai"

• @rosebud_ai, 2025-08-21 (source)

"simulator made on Rosebud AI"

• @rosebud_ai, 2025-07-27 (source)

"Prizes from both @MeshyAI & @Rosebud_AI 6 months of free sub on both Meshy & Rosebud AI (1st place) 4 months of free sub on both Meshy & Rosebud AI (2nd place) 3 months of free sub on both Meshy & Rosebud AI (3rd place) Winners will be announced shortly after the"

• @rosebud_ai, 2025-06-03 (source)

"This Horror Game Was Made With Rosebud AI #horrorgame #gamedev #ai #3dgamedevelopment #rosebudai #gameassets #3dgames"

• @rosebud_ai, 2025-05-22 (source)

Quick Facts

Field	Value
Product	Rosebud AI (text-to-game maker)
URL	rosebud.ai · play.rosebud.ai · lab.rosebud.ai
Category	AI game creation / vibe coding
Founded	2019 (YC S19)
Founders	Lisha Li (CEO 2019-2025), with Carlos M. Cuya and Boris Dayma in the early team
HQ	San Francisco
Team size	~10
Funding raised	$8.8M across seed + Series A
Lead investors	Khosla Ventures (seed), Animoca Brands (Series A), Y Combinator
Series A	$6.6M, December 2021
Reported MRR	~$40K (indie-scale estimate)
Pricing	Free (20 prompts/week), Pro $29/mo, 10x Dev (higher tier, commercial use)
Community	17,358 Discord members, ~75K-person waitlist at peak (Jan 2024)
Tech stack	Next.js front end, Phaser.js runtime for 2D, browser-hosted code editor, LLM agents (Claude/GPT class) for code gen, in-house image models for assets
Current status	Live, growing, founder transitioned to a16z mid-2025

The Three Pivots — What an Indie Hacker Should Actually Learn

This is the part of the Rosebud story worth your attention. Skip the product walkthrough if you have to, but read this section.

Pivot 1 (2019-2021) — Synthetic Media for Fashion Marketing

The original Rosebud was a synthetic media company. The pitch was that brands would soon stop doing photo shoots and instead generate realistic AI faces and bodies wearing their clothing. Lisha had a PhD in machine learning from Berkeley and had spent time at Pinterest and Stitch Fix before becoming a principal at Amplify Partners, so the technical thesis was credible and the customer thesis was logical. They got into YC S19, raised seed money from Khosla Ventures, and built the tooling.

It did not work. Not in the sense of "the tech failed" — the tech was fine — but in the sense that the buyers (marketing teams at fashion brands) were extremely slow to adopt, the customization controls were not yet good enough for production use, and the cost per image was high. The product attracted attention but not money. The TAM of "fashion brands who will replace photo shoots with AI in 2020" turned out to be approximately zero.

What an indie hacker should notice: even an obviously correct long-term thesis (yes, AI will replace photo shoots) can be unmonetizable for years before the customers actually move. If your runway is shorter than the adoption curve, the thesis is wrong for you regardless of whether it is right about the world.

Pivot 2 (2021-2023) — Tokkingheads, Then NFT-Adjacent Character IP

In 2021 the team shipped Tokkingheads, an app that animated a single portrait photo into a talking avatar using voice or text input. This was a real product. Millions of people used it. It went viral on social media in short bursts. But it had no obvious monetization path: the typical user was a teenager turning their selfie into a singing meme, not a business buyer.

Then NFTs happened. In December 2021 Rosebud raised a $6.6M Series A. The lead investor was Animoca Brands, which at that moment was the most aggressive NFT/Web3 gaming investor on earth. The product narrative shifted accordingly. The pitch became something like: AI-generated characters as on-chain IP, animated avatars for the metaverse, programmable NPCs you could trade. Rosebud built tooling for character generation that pointed at this direction.

By mid-2022 the NFT market had collapsed, and by late 2022 the entire generative AI conversation had moved to LLMs (ChatGPT launched November 2022). The "characters as NFTs" pitch was dead inside of twelve months. Rosebud had raised at a peak and now had to find a thesis that worked without the Web3 buyer.

What an indie hacker should notice: raising money from an investor with a strong directional bias (Animoca = Web3) pulls your roadmap in their direction. That is not always a problem. It became a problem here because the bias was wrong. If you take money, take it from someone whose worldview can survive a hype cycle ending.

Pivot 3 (2024-now) — Text-to-Game in the Browser

In January 2024 Lisha posted a Launch HN titled "Rosebud (YC S19) - Turn game descriptions into browser games." The pitch was tight: type what you want, get a playable Phaser-based game in your browser, clone other people's games as starting points, generate sprites and music in-line. They described the product as "ChatGPT + Midjourney + Replit."

This pivot worked for three reasons that compound:

1. LLM code generation finally crossed the quality threshold. The 2021 version of this idea would have produced unusable code. The 2024 version, on top of GPT-4-class models, produced something a non-developer could iterate against.
2. The Tokkingheads user base became free distribution. Hundreds of thousands of people already knew Rosebud as the company that does fun creative AI things. When the game maker shipped, a non-trivial slice of those people tried it.
3. Phaser as a runtime constraint made the AI's job tractable. General "code anything" is too open-ended for an LLM to produce reliably. "Code a 2D Phaser game" is constrained enough that the agent can be reliable in production. This is the under-appreciated technical decision.

Within weeks the waitlist hit 75,000 people. By end of 2024 user-generated projects on the platform were up roughly 300% from launch. The Discord crossed 17K members. Pro subscriptions started compounding. The company became a small but real business.

What an indie hacker should notice: a tight constraint (we only build Phaser 2D games, not "anything") was the technical decision that made the product reliable enough to charge for. Most failed AI agent products are too general.

5-Minute Product Walkthrough

I went to rosebud.ai. The hero is a single text box with a 500-character limit and a Create button. No video, no logos, no testimonial wall. The tagline reads "Make 3D Games & Worlds with Vibe Coding." There is a row of templates below: top-down RPG, platformer, endless runner, narrative visual novel, idle game.

I typed: "a top-down maze where you collect dropped umbrellas while avoiding rain clouds." Sign up, no credit card, free tier active. Click Create.

The screen splits. Left side: a chat-style transcript where I can see the agent narrating what it is building ("Setting up Phaser scene... generating player sprite... wiring collision detection..."). Right side: a live preview window that shows the game compiling in real time. Roughly 60-90 seconds in, an actual playable game appears. Arrow keys move the sprite. Bumping into a cloud kills me. Bumping into an umbrella increments a counter at the top of the screen.

The art is rough. The "umbrella" sprite is a vaguely umbrella-shaped object in a flat color palette. The "rain cloud" is a gray blob. But it works. I can hit a refine button, type "make the umbrellas more colorful and add a 30-second timer," and watch the agent re-edit the code. The second iteration takes maybe 30 seconds. Now I have a timer.

There is a code panel I can open if I want to read what was generated. It is real JavaScript using the Phaser 3 API. I can edit it directly, and the agent respects my edits on subsequent iterations — mostly. (Top complaint in the Launch HN comments was that manual edits sometimes got overwritten. The team acknowledged this and said they were working on diff-based regeneration.)

There is a Publish button. Hitting it gives me a shareable play.rosebud.ai/g/[id] URL. The game runs in any modern browser. There is a community feed where I can see what other people published. Most of it is rough, some of it is genuinely fun.

This is the entire product. It is not deep. It does exactly one thing and it does it competently for someone who has never written code in their life.

What is missing: serious tooling for actual game developers (no asset pipeline beyond what the AI generates, no version control, no real debugger, no mobile export, no in-app monetization for creators). This is intentional. The product is for the hobbyist who wants to ship a thing this weekend, not the studio that wants to ship a thing this year. Trying to serve both would have killed it.

Business Model Deep Dive

The numbers I am about to give you are estimates triangulated from public signals (Pro pricing, Discord size, Crunchbase, founder commentary, comparable indie-scale AI tools). Treat the MRR figure as a rough order of magnitude, not a financial statement.

Pricing structure

• Free — 20 prompts per week. Can create and edit projects. Cannot publish games for commercial use.
• Pro — $29/month. Significantly higher prompt allowance. Commercial use rights. Priority generation.
• 10x Dev — higher-priced tier (not publicly listed, contact-for-pricing) aimed at users who want larger context windows, faster generation, and full commercial rights.

The free tier is generous enough to actually evaluate the product. The Pro tier sits at the standard $20-30 price band that AI consumer tools have converged on (Cursor $20, ChatGPT Plus $20, Midjourney $30). This is the band where you can charge enough to cover inference costs but cheap enough to not trigger a budget approval in the user's head.

MRR estimate

If we assume the Discord (17K members) is roughly 5x the size of the active paying base — a reasonable ratio for a freemium AI tool where a lot of Discord lurkers tried the free tier and never paid — we get something like 3.5K active users. If 35-45% are on a paid plan (high for SaaS, normal for tools where the free tier is gated tightly enough to push power users to upgrade), that is roughly 1,200-1,500 paying customers. At a blended $30 ARPU, that gives a band of roughly $35K-$45K MRR. The $40K midpoint is the number I have used in the title.

For context: that is roughly $480K ARR. With ~10 employees in San Francisco, that does not cover payroll on its own. The business is propped up by the $8.8M they raised, and presumably runway is enough to keep iterating. The MRR is real but the business is not yet self-sustaining at venture-scale headcount.

Why $40K MRR is impressive given the context

Most YC companies that raise $6.6M Series A and pivot twice are dead before their seventh year. Rosebud is not dead. It has a paying customer base in the low thousands, an engaged community in the tens of thousands, and a product that works. For an indie hacker, that is the ceiling of what most realistic AI tool products can hope to reach in 18-24 months. For a venture-backed company, it is a difficult middle place — too big to shut down, too small to IPO. The CEO's move to a16z suggests the conclusion: stabilize, hand off, redeploy the founder's time elsewhere.

Unit economics that matter

The expensive part of this business is LLM inference. Every Create button click runs an agent loop that may make a dozen LLM calls (decide what kind of game, generate scene code, generate sprite, regenerate on error). At Anthropic/OpenAI API prices, a single game generation probably costs Rosebud somewhere between $0.10 and $0.50 in raw inference. The free tier (20 prompts/week) probably loses money per user. The Pro tier breaks even or makes modest margin depending on how heavily the user generates.

This is the structural reason the company needs to be careful with free-tier limits. Every free user is a loss leader. The 20-prompts-per-week cap is not arbitrary — it is the largest number that does not bleed the company.

Tech Stack Reverse-Engineered

There is nothing here that requires a research lab. This is a wrapper around publicly available models, with a smart choice of game framework underneath.

Front end — Next.js, served from rosebud.ai. The IDE-in-browser is built with a Monaco-style code editor on the right pane and a Phaser preview iframe on the left. Standard SaaS plumbing.

Game runtime — Phaser 3, the open-source JavaScript 2D game framework. All 2D games on Rosebud run on Phaser. The team's official Phaser partnership announcement in January 2024 confirms this directly. The 3D side appears to use a Three.js or Babylon.js layer; the public surface is less specific about the 3D tooling.

Code generation — LLM agents calling Claude/GPT-class models (the company has not disclosed which). The agent loop is multi-step: a planner decides what kind of game to build, a code generator produces the Phaser scene, an asset generator produces images, an error handler watches for failures and retries. This is standard 2024-era agent architecture. The proprietary part is the prompt engineering and the constrained code template library that keeps generations on-rails.

Asset generation — In-house image models (originally derived from Tokkingheads-era work) for sprite generation, plus some routing to general-purpose image models for backgrounds. The 3D asset side appears to use third-party services.

Hosting and playback — Published games run as static bundles in the browser. The play.rosebud.ai URLs serve a packaged Phaser scene plus assets via what looks like CDN-delivered static files. No server-side game logic. This is cheap to scale.

Database — Almost certainly Postgres for user accounts, project metadata, and community feed. Standard.

Payments — Stripe, based on the checkout flow.

The point: an indie hacker with $50K, six months, and reasonable Next.js skills can build the structural equivalent. The hard part is not the engineering. It is everything in the next section.

Distribution Playbook — What Actually Drove the Growth

This is where Rosebud's story diverges from a generic indie hacker template and where the lessons are hardest to copy. Three sources of traffic, in rough order of importance:

1. Lisha Li's personal brand inside the AI community

Lisha was a Berkeley ML PhD and a principal at Amplify Partners before founding Rosebud. She is well-known in the West Coast AI scene. She has 20K+ Twitter followers, regularly speaks at AI events, and is part of the social graph that includes a16z, Khosla, OpenAI alumni, and senior researchers. When she posts that Rosebud shipped a new feature, the people who reshare it are not random — they are influencers in the audience that includes Rosebud's target user (AI-curious hobbyists and indie developers).

You cannot copy this directly. You can copy the structural insight: in a noisy AI tools market, the founder's personal authority in their niche is the cheapest distribution channel that exists. If you do not have it, build it for 12 months before you launch. Post technical breakdowns, ship side projects with public code, become known for one specific opinion.

2. Launch HN and Show HN

The January 4, 2024 Launch HN post was the single biggest distribution event in Rosebud's recent history. It put the product in front of the technical-curiosity audience at exactly the moment that audience was primed to try AI agent tooling. The post drove the waitlist to 75K within weeks. The Phaser partnership announcement followed days later and amplified the loop.

You can copy this. Plan your Show HN months in advance. Make sure the product is interesting enough to survive HN's brutal commentary, which is itself the marketing — every critical comment that the founder responds to thoughtfully becomes social proof.

3. Indie game community cross-pollination

Rosebud has an active itch.io presence, runs game jams with the Phaser community, and operates a Discord that functions as both support channel and showcase. The flywheel is: people build games, post them in the Discord, the games are shared on Twitter, that drives signups, which produces more games.

The under-appreciated part: Rosebud actively reposts user games on their own Twitter and gives community members visible status (showcase channels, Discord roles for power users, weekly themed challenges). This is the cheapest user-generated-content marketing engine that exists, and most AI tool companies fail to build it.

What is conspicuously missing

No SEO play. No paid acquisition that I can detect. No influencer marketing beyond the founder. No app store presence. The growth is entirely organic and community-driven. This is sustainable at small scale but is a ceiling-setter — the company is not going to 10x from here without a paid acquisition motor, and that motor would have to work on top of unit economics that are already thin.

Why This Works, Why Now

Three things came together in late 2023 to make this product possible. None of them existed in 2019 when Lisha started the company.

LLM code generation crossed the threshold. Before GPT-4, AI-generated JavaScript was 60-70% correct, which is not good enough to ship a playable game without a developer in the loop. After GPT-4 and Claude 3, on a constrained domain like Phaser scene generation, the success rate is closer to 90-95%. The remaining 5-10% the agent handles by retrying. This is a fundamental capability change that unlocked the entire product category, not just Rosebud.

The "vibe coding" cultural moment. The idea that you can describe software in natural language and get working code is suddenly a mainstream concept, partly through Bolt, Lovable, v0, Cursor, Replit Agent. Rosebud benefits from this rising tide. The phrase "vibe coding" appears in their tagline. Users arrive already understanding what to expect, which compresses the onboarding.

Browser as the universal game platform. No app store. No install. No platform tax. A published Rosebud game runs in a tweet, a Discord message, a school computer, a phone. The implicit competitor (Roblox) requires an account and a download. Rosebud games require neither. This is the right unit of distribution for a hobbyist's first game.

The window is open right now and probably stays open for another 18-24 months. The risks: a general-purpose tool like Bolt could add a "make me a game" template and erase 70% of Rosebud's moat overnight. The defense: vertical specialization, community, and the fact that game-shaped problems are different enough from app-shaped problems that the general tools will be mediocre at them for a while.

Founder Profile — Lisha Li

Lisha Li grew up in Toronto. She was a unionized teen actress while in high school, which is the kind of biographical detail that signals comfort with performance and audience-facing work — a useful trait for a consumer AI founder.

Undergrad in pure mathematics. PhD in machine learning from UC Berkeley (2017). Postdoctoral work in deep learning. By the time she finished her PhD, the field had moved from "interesting research" to "this will be the platform of the next decade," and she made the call to leave academia.

She did data science at Pinterest and Stitch Fix — both companies where ML was actually deployed into production at consumer scale. Then she became a principal at Amplify Partners, a respected early-stage venture firm, where she led investments in AI startups. Roughly three years of investing taught her enough about the operator side to want to build rather than fund.

She founded Rosebud in 2019 without a co-founder, which is unusual and which she has spoken about: she had the thesis and did not want to wait. The team grew to roughly 10 over the next several years. She raised $8.8M across rounds without selling the company at any point. She survived two pivots that would have killed most founders.

In mid-2025 she announced she was stepping down as CEO and joining a16z as an investing partner on the infrastructure team, focused on the frontier model stack. Rosebud continues under new leadership.

What the founder profile reveals about the company: this is a founder with technical credibility, investor relationships, and personal-brand reach in the AI community. She got the company to a sustainable but small outcome and then redeployed her time to where the leverage was higher — back to investing, but now with operator scars. That is a rational career move and it tells you something honest about Rosebud's likely ceiling.