Bolt.new: What It Builds and Where It Stops

Bolt.new, usually just called Bolt, is an AI builder for websites, web apps and mobile apps. You describe what you want in a chat, and Bolt writes the code, runs it in a live preview, adds a database when the app needs one, and publishes it to a web address, all inside your browser with nothing to install. It is made by StackBlitz and runs on StackBlitz’s WebContainers. It builds JavaScript apps: a Node.js backend, any front-end framework, and mobile apps through Expo. It stops where JavaScript stops, where its context window fills up on a large project, and where you want the code to live in your own repository with your own tests and review. That is the point to hand over to a coding agent.

How Bolt works

Bolt’s editor, terminal and preview all run in the tab. StackBlitz describes WebContainers↗ (opens in a new tab) as “a browser-based runtime for executing Node.js applications and operating system commands, entirely inside your browser tab.” That is why Bolt needs no local setup and behaves the same on Windows or a Mac, and it is also why it works best in Chrome and other Chromium-based browsers. In Safari, Bolt’s own help center says the preview is hosted rather than run in the browser and the code view is read-only.

The AI side is an agent that plans, writes code and fixes errors. You pick one of two: Standard, fast and suited to well-defined tasks, or Max, which reasons more and suits large codebases, refactors and open-ended problems; Max needs a paid plan. Bolt chooses the model behind each agent and may change it as new models arrive, so you pick a level of effort, not a model. Plan Mode lets you discuss an approach, debug or research without changing any code, then turn the plan into a build.

What it builds

• Websites: landing pages, company sites, portfolios and small shops.
• Web apps: dashboards, booking tools, CRMs, internal tools and SaaS prototypes, built on JavaScript frameworks.
• Mobile apps: through the Expo integration. Put the words “mobile app” in your first prompt so Bolt sets up an Expo project from the start.
• Starting points other than a prompt: a Figma frame, a Google Stitch design, a GitHub repository, or a project imported from Lovable.

Can Bolt make a backend? Yes, in JavaScript. Bolt’s supported technologies page↗ (opens in a new tab) says it supports Node.js for the backend and only JavaScript-based backends, so PHP and Python are not compatible. Bolt Cloud supplies the rest: a database Bolt creates automatically when your app needs one, with tables, authentication and user management, file storage, secrets, logs and server functions for work that should not run in the browser, such as payments or webhooks. On some plans you can use Supabase instead. Sign-up and login are not always added automatically; ask for them.

How to use Bolt.new well

1. Start in Plan Mode for anything larger than a page. Agree the screens, data and order of work before Bolt writes code.
2. Write the first prompt in full, then keep later prompts short and focused on one change. One prompt asking for many changes is the usual reason Bolt builds only part of it.
3. Put standing instructions in project knowledge or an agents.md file, which Bolt finds and uses automatically, instead of repeating them in every prompt.
4. Check the preview after each change, and name or bookmark good versions in version history so you can get back to them.
5. Clear the context with /clear when you finish a feature. Bolt still sees the code, but no longer carries the old chat.
6. Run the security audit from the Publish menu before every publish, then publish.

Export and deploy

Publishing is built in. Every project can go live on a free bolt.host address, and paid plans can connect a custom domain; you can also publish through Netlify instead of Bolt hosting. Publishing does not update your live site automatically after later edits: click Publish again to push changes.

Getting the code out takes two forms. Export → Download gives you a zip of the project files at any time. The GitHub integration keeps a repository in step: Bolt commits each change that does not break the project and checks GitHub every 30 seconds for commits made elsewhere, while branches can be created in Bolt but must be merged on GitHub. Bolt says the code you create is yours to use, including commercially. Moving that repository into a coding agent’s hands, and the rare case where Bolt and GitHub change at the same moment, are covered in Lovable, Bolt and v0 vs Claude Code.

Where Bolt stops

• JavaScript only: no Python, PHP, Go or Java backend. If your stack is one of those, Bolt is the wrong tool.
• Context: Bolt keeps recent messages in its working memory, and its guide to managing context↗ (opens in a new tab) says a growing project can bring errors, slower replies and weaker results. It suggests clearing context, removing unused files, and splitting files over about 700 lines.
• Undo does not cover data: restoring an earlier version from version history does not change your Bolt or Supabase database.
• No outside control: Bolt’s FAQ says it has no public API, CLI or SDK, so you cannot script it from CI or another agent.
• Idle databases: in unpublished projects, a database with low use for six days or more may be paused, and restored when you next open the project.
• Mobile browsers: you can prompt and publish from a phone, but code view, version history, visual edits and the database need a desktop.
• Security is checked, not guaranteed: the audit runs on paid plans and a lighter database check runs on all plans. Neither replaces your own review; see vibe coding security.

When to move to a coding agent

Bolt and a coding agent such as Claude Code, Codex or Cursor are not rivals so much as stages. Bolt is fastest from nothing to a working, hosted first version. A coding agent works in a repository you control, runs your tests and fits your review process. Signs you have reached the handover:

• Prompts start undoing earlier work, or Bolt struggles to hold the whole project in view.
• You need a language or service Bolt does not support, or a build step and test suite that run in CI.
• More than one developer is changing the code and you want pull requests and code review.
• The app now holds real customers’ data and needs a deliberate security review before each release.

Connect GitHub first, stop editing in Bolt or give it its own branch, and let the agent work through pull requests. An agents.md written for Bolt carries straight over, because most coding agents read AGENTS.md. For where builders sit on the wider map, read vibe coding vs low-code; for the closest alternative builder, Replit Agent.

Keeping the work list outside the chat

Bolt’s chat is a good place to build and a poor place to remember what is left to do, especially once you clear context or hand the project to someone else. fenbs keeps that list on a board: each feature, enhancement or bug is a task with a note, a plan and a test status, in lanes To Do, Next Up, In Progress and Completed. Bolt can add a custom MCP server by URL with an MCP OAuth sign-in, which is how fenbs’s server at https://fenbs.ai/api/mcp signs clients in, so you can add the board as a connector; Bolt switches on every tool a connector offers, so turn off the ones you do not want it to use. When the project moves to a coding agent, the same tasks go with it, and History shows who changed each one. fenbs has no due dates or sprints; it is a list of work people and assistants share.