AI code generation (codeGen) is quietly becoming one of the most important unlocks in crypto.

It’s already a major wave in web2. Tools like Cursor, Lovable, and Replit are giving anyone the power to go from idea to app, often in minutes. These tools didn’t even exist a few quarters ago. Now they’re foundational.

We’ve been developing an internal thesis on how AI and crypto intersect, and why codegen will reshape the way consumer crypto products are built.

This post walks through:

• A quick primer on the web2 codegen landscape

• Why building crypto apps still hits some unique friction

• The web3 codegen players worth watching

• A small experiment building a dApp across tools

• And where we think all this is going next

Foundational AI models have been the focus of most hype, funding, and adoption. But over the past year, code generation has become one of their most useful and consistent applications.

Just look at the SWE-benchmark scores: Claude has jumped from 18% to 70% in a year. Benchmarks aren’t perfect, and devs have their favorites, but the takeaway is clear — codegen is now good enough. With solid product sense and a systems mindset, anyone can ship an MVP:

These models are getting very good at code generation over the past year. Every couple of months, the leaderboard changes, Anthropic is leading the way. As a point of reference, Claude has improved its SWE-benchmark for coding issues from 18% to 70% YoY. Benchmarks are a way of measuring, anecdotally, developers may have different stories on which model is more performant, but the bottom line is the same: It’s now possible. Any person with a good systems design and first principles product brain is able to build an MVP.

To understand the codegen space, it helps to think across two axes:

1. Product lifecycle 

• Ideation: Brainstorming, outlining, and stress-testing product ideas

• Design: Turning ideas into wireframes or frontend components

• Frontend: UI logic and interactivity

• Backend: Data and execution layers

• Testing: Ensuring code actually works

2. Level of assistance

• Copilot: e.g., GitHub Copilot, Claude, ChatGPT prompts. Made for devs.

• Assisted IDE: Cursor, Windsurf. Great for beginners or semi-technical users.

• Cloud IDE / App Builder: Bolt, Lovable, Replit. For non-devs; from prompt to deployable app.

• No-code: Softr, Bubble, Glide, mobile app generators. Abstract away almost everything.

Here’s a rough mapping of tools across the lifecycle and skill levels:

• Cursor: MIT founders, launched 2023. Funded by a16z and OpenAI. Targets serious devs. Reached ~10b val.

• Windsurf: Ex-Codeium, MIT team. $150M led by General Catalyst, could be bought by OpenAI for $3b. Deep IDE integration.

• Lovable: Viral Swedish startup with ~$50M ARR. Built by AI researchers, for non-devs.

• Bolt: Spun out of StackBlitz. Also reached ~$50M ARR. Built for non-devs with real deployment capabilities.

Integrations: Cursor/Windsurf plug into developer workflows (VS Code, GitHub, etc.). Lovable and Bolt tie into web services: Lovable connects to Supabase and GitHub, while Bolt integrates with common stacks (e.g. its examples use Supabase).

Monetization: All offer a free tier to drive adoption, then upsell through either flat-rate SaaS subscriptions (Cursor, Windsurf, Replit) or usage-based pricing models (Bolt, Lovable), at the enterprise level, it’s a mix of seat + usage based pricing. Underneath, they all rely on foundation models like GPT or Claude, and each is building a wedge around either workflow integration, proprietary models, or control over hosting and deployment.

There are two typical paths:

Hard Mode (for technical builders):

• ChatGPT: Prompt refinement, ideation, technical research

• Builder: Design + frontend code (auto-push to GitHub)

• Cursor / Windsurf: Backend logic, DB management, testing

Easy Mode (for non-devs):

• ChatGPT: Same starting point

• Lovable / Bolt: Full-stack prompt-to-app generation

Peter Yang has a great newsletter doing walkthroughs using this stack.

Here’s a couple of examples of apps used with this stack:

So what’s different when building crypto apps?

Most of the frontend and some backend logic can be scaffolded using the same tools. But the moment you touch on-chain logic or integrations, things break.

Main roadblocks:

• Wallets and smart accounts: Codegen tools don’t yet support Privy, Turnkey, Rhinestone, Magic, Fireblocks, etc.

• Chain selection and RPCs: Which L1/L2 to build on? Which toolchain fits best?

• On/Offramps: No easy way to add crypto ramps like Stripe or Coinbase Pay.

• Security and upgradeability: Would you trust AI-generated code with real user funds?

• Monetization: Web3 models are fundamentally different (fees on txs, app tokens, etc.

We’ve used and been in touch with a couple of teams building in the space:

Ohara and Dev are leading the way in sheer terms of users and apps being built, but their focus on mostly around app tokens, and even though it’s a short term profitable strategy, a lot of these apps and users end up in a dilemma with tokenholders and their rights. 

Ohara is the only native launchpad, as Dev leverages pump. On Ohara, every app creator gets 10% of the supply vested over 12mo, and every app remixed gets 3% of the supply, and Ohara’s $HELLOWORLD token will capture 1% of the token supply of all tokens launched on Ohara.

As the web2 market shows, it’s a matter of being great at a specific lane, and solving the real problems that users are facing. Right now, no one is focused on the later stages and copilot assistance for web3 apps. I’d be eager to see a platform that has deep native integrations that solve account abstraction, key management, on/off ramps, miniapp distribution and smart contract code. Rather than try to displace web2 incumbents, work as extensions.

We tested building a basic crypto portfolio tracker across several platforms, using Coingecko’s API:

Most tools can handle the frontend, but compute limits and real-time API access remain major blockers. Ohara performed best, though still had its quirks.

One thesis we’re watching closely: social creation.

Imagine prompting a miniapp from inside your TikTok feed. Not a filter, a working app, generated on the spot, tied to your wallet, and remixable by your followers.

We think this format, ephemeral, personalized, and viral, is inevitable.

Expect to see the first experiments soon (e.g., Vibecode).

We’re bullish on web3 codegen. Whether you’re:

• A builder looking to ship faster using crypto primitives

• A creator spinning up ephemeral miniapps for your audience

This is where the next wave of crypto UX is heading. 

Codegen has raised the bar. Pitch decks aren’t enough anymore, shipping is the new demo.

This post is for informational purposes only, and does not constitute a recommendation to buy or sell securities or to pursue any particular investment strategy. This post should not be relied upon in evaluating the merits of any investment or any particular investment strategy. You should consult your own advisers as to business, financial, tax, legal, and all other related matters concerning any investment. The views expressed in this post reflect the current opinions of the authors and do not necessarily represent the opinions of Social Graph Ventures LLC