Bolt.new for Vibe Coding: WebContainers, Claude, and $40M ARR in 6 Months

What Bolt.new Actually Is

Bolt.new is a browser-based development environment that turns text prompts into deployed web applications. You type what you want. An AI agent writes the code, installs npm packages, starts a dev server, and shows you the running app, all inside a browser tab. No local setup. No terminal. No git clone.

The product launched in October 2024 as a pivot from StackBlitz, a company that had spent years building WebContainers, a technology for running Node.js inside web browsers. Before Bolt, StackBlitz was a developer-tools company with modest traction. After Bolt, it reached $4M ARR in 30 days, $20M in 2 months, and $40M ARR by March 2025. That trajectory made it the second-fastest SaaS product to reach $40M ARR in history, behind only ChatGPT.

StackBlitz raised a $105.5M Series B in January 2025 led by Emergence Capital and GV (Google Ventures), with participation from Madrona and Conviction. The company was valued at $700M, roughly 17.5x its ARR at the time.

The term "vibe coding" was coined by Andrej Karpathy in February 2025 to describe a style of programming where you "fully give in to the vibes, embrace exponentials, and forget that the code even exists." Bolt.new is the canonical vibe coding tool: you describe what you want, the AI builds it, and you interact with the running result rather than reviewing diffs.

How WebContainers Work

WebContainers are the technical foundation that makes Bolt possible, and the reason it feels fundamentally different from tools that generate code and hand it to you.

A WebContainer is a WebAssembly-based operating system that runs inside a browser tab. The Node.js runtime, originally written in C++, is cross-compiled to WASM, which lets it execute natively inside the browser's JavaScript engine at near-native speed. A dedicated Service Worker handles the virtualization layer, intercepting file system operations when Node.js processes read or write files. A virtual file system manages all project files, source code, configs, and dependencies within the browser.

The result: when you prompt Bolt to build a Next.js app, it does not send your code to a remote server. It boots a Node.js process in your browser tab, runs npm install, starts the dev server, and renders the live preview right next to your code. Latency is zero because nothing leaves your machine.

WebContainers run in all major browsers (Chrome, Firefox, Safari) and work offline once loaded. This is a genuine architectural difference from cloud-IDE tools like Replit or GitHub Codespaces, which spin up remote VMs and stream the results back over a network connection.

The AI Engine: Claude Models in Bolt

Bolt started with Anthropic's Claude 3.5 Sonnet as its primary code generation model. As of April 2026, it offers the full Claude model family with the ability to switch models per conversation.

Sonnet 4.5 is the default and recommended starting point. Bolt's documentation describes Sonnet 4.6 as offering "Opus-class capability at a pace and price built for all-day use," making it the practical sweet spot for complex work. Opus 4.5 and 4.6 are reserved for problems where accuracy matters more than speed.

The Claude Agent in Bolt is distinct from the legacy v1 Agent. The Claude Agent supports Plan Mode (thinking before coding), creates Bolt Databases, connects to Supabase, and produces more architecturally consistent output. The v1 Agent is being retired: unavailable for new projects after April 13, 2026, and fully inaccessible after August 3, 2026.

What You Can Build

Bolt supports React, Next.js, Vue, Svelte, Astro, and Vite on the frontend. Styling options include Tailwind CSS, CSS Modules, Styled Components, and Sass. On the backend, you get Node.js with Express or similar frameworks, plus Bolt Database or Supabase for persistence.

The practical ceiling is complexity. Bolt handles single-developer prototypes and small apps well. Once you exceed a few screens with interconnected backend logic, the generated code gets harder to maintain. Multiple reviewers report that Bolt produces flat, inconsistent architectures that work but do not scale. For production-bound projects, the common pattern is: prototype in Bolt, then migrate the code to a proper IDE for hardening.

Bolt v2 and Plan Mode

Bolt v2 launched in October 2025 with three changes that addressed the platform's biggest pain points.

Autonomous debugging. Bolt v1 was notorious for error loops: the AI would hit an error, attempt a fix, break something else, attempt another fix, and burn through tokens in a spiral. Bolt v2's debugging system reduces these loops by 98% according to StackBlitz. The agent now reads terminal output, identifies root causes, and applies targeted fixes instead of regenerating entire files.

Plan Mode. Before writing code, the Claude Agent can produce a structured plan: what it intends to build, which files it will create, what dependencies it needs. You review and adjust the plan before any tokens are spent on code generation. This is particularly valuable for complex prompts where ambiguity leads to wasted iterations.

Built-in infrastructure. Bolt v2 added Bolt Database (powered by Supabase), integrated hosting, authentication, and payment processing. New projects get a complete backend stack without manual configuration. You can also connect to an existing Supabase project, use the same database across multiple apps, and pause/resume database instances from within Bolt.

Pricing Breakdown

Bolt uses a token-based pricing model. Every AI interaction consumes tokens. The more complex the prompt and response, the more tokens it costs. On paid plans, unused tokens roll over for one additional month.

The free plan gives you roughly 10-20 meaningful interactions for a medium-complexity project. That is enough to test the platform, not enough to build anything real. The Pro plan at $25/month is the practical entry point for actual development.

Annual billing saves up to 28% across all paid plans. Token rollover (effective since July 2025) means unused Pro tokens carry into the next month, giving you up to 20M tokens in a billing cycle if you have a light month.

Strengths

1. Zero Setup, Instant Preview

No installation. No environment configuration. No dependency management. Open bolt.new in a browser, type a prompt, and see a running app in seconds. For developers who want to test an idea before committing to a project, this friction removal is significant. For non-developers, it is the difference between building something and not.

2. The Live Dev Server Matters

Most AI code generators produce files. Bolt produces a running application. The WebContainer boots a real Node.js server in your browser, so you interact with actual HTTP responses, working forms, real routing, and live database queries. This is not a static preview. It is the app, running.

3. Full Code Ownership

Unlike managed platforms where backend code is locked in, Bolt gives you full access to every file it generates. You can download the project, push it to GitHub, and continue development in any IDE. There is no vendor lock-in on the code itself.

4. Framework Breadth

React, Next.js, Vue, Svelte, Astro, Vite, Expo. Bolt supports the major JS frameworks and lets you pick the right tool for the project. Most competitors are locked to a single framework (Lovable: React, v0: React/Next.js).

Limitations

1. JavaScript Only

This is the constraint that defines what Bolt can and cannot be. WebContainers compile Node.js to WASM. Python, Go, Ruby, PHP, Java, Rust: none of these run in a WebContainer. If your backend needs any non-JS language, Bolt is not an option. This rules out Django, Flask, FastAPI, Rails, Spring Boot, Gin, and every other non-JS server framework.

2. Code Quality Degrades with Complexity

Bolt-generated code works for prototypes. Architecturally, it is flat. Patterns are inconsistent. Files get large. Once your app exceeds a few screens with interconnected state and backend logic, you spend more time fighting the generated code than building new features. This is the experience reported by developers across multiple independent reviews.

3. Debugging Eats Tokens

Despite v2's improvements, complex debugging sessions still consume large token budgets. The AI sometimes rewrites entire files to fix a single bug, breaking UI structure while failing to resolve the original problem. A bad debugging session on the Pro plan can burn a significant chunk of your monthly 10M tokens.

4. No Human Support

Support is handled entirely by AI and community forums. There is no human support team to escalate to. For a paid product at $25+/month, this is a gap. When you hit a platform-level issue (not a code issue), there is no path to resolution beyond waiting.

5. Browser-Only Workflow

Everything happens in the browser. You cannot use your local editor, your terminal aliases, your debugging tools, or your git workflow. For developers with established environments, this is a downgrade in ergonomics. For non-developers, it is the whole point.

Bolt vs Alternatives

Lovable generates production-grade TypeScript with Supabase and handles deployment. Better for apps that need databases, authentication, and payments from the start. Lovable hit $300M ARR in January 2026, making it the largest player in the category.

Cursor is an AI-native desktop IDE. It puts you closest to the code: you write it, review it, and deploy it, with AI accelerating each step. Better for professional developers working on complex, long-lived codebases. Different category entirely.

Replit runs a full cloud VM with support for any language. If you need Python, Go, or a multi-language stack, Replit handles it where Bolt cannot. Slower startup, broader capability.

v0 by Vercel generates clean React and Next.js components from prompts. It does not provide a runtime or backend. Best for generating UI code you will integrate into an existing project.

For a full comparison, see our guide to the best AI app builders in 2026.

WebContainers vs Server-Side Sandboxes

Bolt's WebContainers and server-side sandboxes represent two fundamentally different approaches to running AI-generated code. Each makes trade-offs that determine what you can build.

WebContainers win on latency and simplicity. There is no server to provision, no cold start to wait for, and no network dependency after initial load. For JavaScript-only projects, the developer experience is fast and seamless.

Server-side sandboxes win on versatility. They can run Python ML pipelines, Go microservices, Rust compilers, multi-container Docker setups, GPU workloads, and anything else a real server can handle. The trade-off is network latency and infrastructure cost.

For AI coding tools specifically, the choice shapes what the tool can do. Bolt's browser-based approach means zero infrastructure cost per user but limits the entire platform to the JavaScript ecosystem. Tools built on server-side execution (like Morph's infrastructure) can process code edits across any language at 10,500+ tokens per second, because the compute happens on dedicated hardware rather than inside a browser tab.

Neither approach is universally better. The question is what you are building. A React prototype with a Supabase backend? WebContainers are perfect. A Python API with data processing, or a multi-language project that needs real server resources? You need a server-side sandbox.

Frequently Asked Questions

What is Bolt.new?

Bolt.new is an AI-powered development platform from StackBlitz that runs a full Node.js environment inside your browser using WebContainers (WebAssembly-based OS). Describe an app in plain English and Bolt generates the code, installs dependencies, starts a dev server, and deploys the result. It reached $40M ARR within 5 months of its October 2024 launch, making it the second-fastest SaaS product to that milestone after ChatGPT.

Can Bolt.new run Python backends?

No. WebContainers compile Node.js to WebAssembly for in-browser execution. Python, PHP, Go, Ruby, and other non-JavaScript runtimes are not supported. This is an architectural limitation of the underlying technology, not a product roadmap gap. If your project requires a Python backend, consider Replit (cloud VM with multi-language support) or a local development environment with an AI coding assistant like Cursor.

How much does Bolt.new cost?

Free: 1M tokens/month (300K daily cap). Pro: $25/month with 10M tokens and no daily limit. Teams: $30/month per member. Enterprise: custom pricing. Annual billing saves up to 28%. Unused tokens roll over for one additional month on paid plans.

What AI models does Bolt.new use?

Bolt offers Anthropic's Claude family: Haiku 4.5 (fast, simple tasks), Sonnet 4.5 (default, balanced), Sonnet 4.6 (complex refactors), Opus 4.5 (architectural decisions), and Opus 4.6 (hardest problems). The Claude Agent also supports Plan Mode for thinking before coding.

Is Bolt.new good for production apps?

For prototypes and MVPs, yes. Bolt compresses the timeline from idea to deployed app from weeks to hours. For production at scale, the generated code typically needs hardening. The common workflow: prototype in Bolt, export to GitHub, then refine in Cursor or a traditional IDE. Bolt's Netlify integration supports custom domains, SEO optimization, and analytics on paid plans.

What is the difference between Bolt.new and Lovable?

Bolt runs a live Node.js dev server in your browser (via WebContainers). You see and edit real running code. Lovable is a conversation-first builder that generates TypeScript + Supabase apps and handles deployment. Bolt gives more developer control and framework choice. Lovable produces more maintainable code for full-stack apps and has deeper database/auth integration out of the box. Bolt works for JS-only projects. Lovable targets apps with real backends and user management.