Bolt is the tool you reach for when you want more than a black box. Built on StackBlitz’s WebContainers technology, it generates a complete full-stack project, runs it live in your browser, and — crucially — lets you see and touch every file it creates. That last part is what separates Bolt from the field.

Whether that’s a feature or a burden depends entirely on who you are.

New in May 2026: Image editing, GitHub flexibility, and team admin controls

Bolt’s recent release notes bring four changes that mostly polish rough edges around the v2 launch from last month, but one of them is genuinely useful day-to-day. In-chatbox image editing using the Nano Banana model is the headline. Select an existing image in your project and ask Bolt to change a specific part of it — swap a background, recolor a button, redraw a logo — and the rest of the image stays intact. This is on paid plans only. It’s a small feature in description, but it closes the loop on Bolt’s image-generation story: you can now create, edit, and ship images without leaving the build session.

Three other changes are less exciting but worth knowing. You can now disconnect a single project from its linked GitHub repo without breaking your overall GitHub integration — useful if you forked a project and want to push the new copy to a different repo. Teams admins can now lock site visibility (public-only, private-only, or user-specified) across all team projects, which removes a class of “I accidentally published the prototype” mistakes that small teams have been hitting. And if a published project’s SSL certificate fails to generate automatically, you can now manually trigger a retry from the publish menu rather than waiting for the system to retry on its own schedule.

None of this changes the Bolt vs. Lovable vs. Replit decision in either direction. But the v2 platform from last month is now more livable to actually work in.

New in April 2026: Bolt v2, Claude Agent as default, and a clearer upgrade path

Bolt shipped its largest change in a year this month. Starting April 13, you can no longer create new projects on the legacy v1 agent — every new project runs on Claude Agent (the new v2), which Bolt describes as faster, better at keeping large codebases coherent, and able to pick up the newest Anthropic models as they land. Existing v1 projects continue to work until August 3, 2026, at which point they’ll be auto-migrated or retired. If you have live v1 work, switch it over now rather than at the deadline — your files and sites carry over, but your chat history does not.

A cluster of related changes came with the transition. Sonnet 4.6 is now the default model, replacing the Sonnet/Opus 4.5 pair. The older models have been retired entirely. The practical implication for founders: builds feel noticeably faster, and the model’s ability to hold a multi-file mental map across iterations is better. Credit burn on typical prompts is moderately lower than on v1 at equivalent complexity.

Image generation is now built in. You can describe an image in the chatbox and Bolt produces one with transparent-background support and automatic WebP conversion. Turn it on from Personal Settings → Add-on features. This removes one of the most common reasons people left Bolt mid-session to hop to Midjourney or DALL·E. MCP server support also landed — you can now connect external context sources (Notion, Jira, Linear, databases) directly into Bolt’s working context, matching a feature Lovable shipped earlier. And a smaller but useful change: the Select tool picker now lets you drill through overlapping layers so you can target the actual card instead of the button inside it.

Netted out: v2 is a real upgrade, not a rebrand. If you used Bolt and bounced, it’s worth another pass.

What Bolt actually builds

Start with a prompt, and Bolt scaffolds a complete project: frontend, backend, package dependencies, configuration files. It runs in a real Node.js environment inside your browser tab — no local setup, no Docker, no “works on my machine” nonsense. The live preview updates as it builds. You can see the file tree, open any file in a full code editor, and make direct edits at any point.

This is genuine full-stack output. Bolt will generate a React frontend talking to an Express or Hono backend, wire up a database connection, configure environment variables, and handle the package install. For prototype-to-working-app speed, it’s legitimately fast.

Framework support

Bolt’s framework support is broader than any competitor: React, Next.js, Vue, Svelte, Remix, Astro, Vanilla JS. If your team already has a stack preference, Bolt will honor it. If you want to ship a Next.js app that a developer can later maintain in a standard repo, Bolt generates exactly that.

This flexibility matters for teams with existing technical opinions. Lovable generates React/Supabase every time. Bolt gives you actual choices.

The credit system: where the frustration lives

Here’s where I have to be direct with you: Bolt’s pricing model is its biggest problem.

The free tier gives you a token budget that evaporates faster than you’d expect. A moderately complex prompt — “build me a SaaS dashboard with auth and a database” — can consume most of a free session. Bolt charges per AI generation, and complex prompts require large generations. You will hit the limit mid-project if you’re not paying.

The $20/month plan substantially increases your credit allowance, but heavy users report needing the higher tiers. There’s a usage meter in the UI, but it doesn’t prevent the specific frustration of watching credits drain right when you’re close to something working.

Compare this to Lovable, which offers unlimited-ish messaging on the Pro plan. For iterating on a complex build, Lovable’s pricing is simply more predictable. Bolt is the right choice when you want code ownership and framework flexibility — but budget for it realistically.

The debugging experience: honest assessment

When Bolt’s AI generates broken code, you will know about it. TypeScript errors, failed npm installs, runtime exceptions — they surface directly in the terminal panel. This is not Bolt being bad at its job. It’s Bolt being honest about what code actually looks like.

For developers and technical founders, this is fine. You paste the error into the chat, the AI usually identifies and fixes the root cause, and you move on. The fix rate on common errors is high. The transparency is actually a trust signal — Bolt shows you what went wrong instead of papering over it with a plausible-looking but broken component.

For non-technical founders: the same transparency that developers appreciate will feel overwhelming here. You need at least a baseline familiarity with what an npm error or a TypeScript complaint looks like to work productively with Bolt. Not the ability to fix it yourself — just enough context to describe it coherently. If that sounds like you, great. If it sounds stressful, use Lovable.

The “context window” problem

Bolt’s AI, like all LLMs in a code context, has a limited memory of the project. On large codebases — multiple feature areas, complex data models, significant CSS — it can lose track of earlier work and generate changes that break things it previously built correctly. This is a known limitation, not a Bolt-specific bug, but it’s more noticeable here than in tools that abstract the codebase away from you.

The workaround: keep sessions focused. Build one coherent feature area per session. Don’t try to build an entire app in one multi-hour prompt marathon.

How it compares to Lovable

The honest comparison:

Bolt wins on: Code ownership, framework flexibility, developer handoff, debugging transparency, and the ability to escape the AI when you need to.

Lovable wins on: Non-coder accessibility, Supabase backend integration, iteration smoothness, and predictable pricing for heavy users.

They’re not competing for the same customer. If you’re a technical founder who will eventually hand this code to a developer — or who will maintain it yourself — Bolt’s output is cleaner for that workflow. If you need to build a working SaaS without any technical confidence, Lovable manages the complexity for you.

Deployment

Bolt includes one-click deployment to Netlify. It’s fast and it works. For a prototype or early validation, this is entirely adequate. For production, you’ll want to move the generated project into a proper repo and set up your own pipeline — which is straightforward given that Bolt generates standard, clean code.

What people have built with it

GrepJob — A production job board scraping verified engineering roles from company career pages, with AI filters for salary and tech stack; built by a solo indie developer with a working subscription tier.
Prilo — An AI study platform for competitive student programs (DECA, FBLA, HOSA) with personalized practice; built and deployed as a live product.

Browse more at Made with Bolt.

The verdict

Bolt is the best AI builder for founders and developers who want to own what they build. The code it generates is legible, follows conventions, and uses frameworks that any developer on the market can maintain. The file access and editor make it uniquely trustworthy: you’re not dependent on the AI’s good behavior, because you can intervene directly.

The credit system is a genuine frustration and the non-coder experience is harder than competitors. But if you’ve ever pushed code, read a stack trace, or made a decision about which JavaScript framework to use — Bolt is the tool that will treat you like an adult.

Rating: 4/5 for developers and technical founders. 2/5 for pure non-coders. The 3 is an average that doesn’t serve either group well — know which one you are.

Was this helpful?