Bolt.new Review 2026: StackBlitz's Browser-Native AI Builder Just Added a Third, Much Cheaper Agent
Bolt.new is still the fastest on-ramp from a plain-English prompt to a running, deployable full-stack app, and it still does something no competitor does the same way: it runs your actual Node.js app inside your browser tab via WebContainers, not on a remote VM. What's changed since most people last looked: Bolt shipped a third agent option two weeks before this review -- Bolt Forge, running open-weight models instead of Claude -- on top of a Microsoft Azure/365 partnership, an AWS Marketplace listing, and a real, documented pattern of token-burn complaints that a rough Trustpilot record makes hard to ignore. This review covers what's verified true today, checked live against bolt.new's own site, GitHub, and release notes.
| Core tech (WebContainers) | Genuinely unique |
| Free tier | Real -- 1M tokens/mo, 300K/day |
| Cheapest paid plan | $25/mo (Pro, 10M tokens) |
| Newest agent | Bolt Forge (open-weight, launched Sep 14, 2026) |
| Trustpilot record | 1.5/5, 202 reviews -- a real trust gap |
Bolt.new's technology is the real deal. Its trust record is the part I'd make you read before you pay.
"WebContainers is the one piece of this product nobody has cleanly copied: a WebAssembly-based operating system that runs actual Node.js inside your browser tab, so when Bolt says it 'ran your app,' it means your machine ran it, not a server queue you're waiting on. That's four-plus years of StackBlitz engineering before AI coding was even a category, and it's why Bolt still feels faster than most rivals in the first sixty seconds. But two things sit next to that genuine strength: a token-based billing model where a documented, recurring pattern of users say Bolt's own bug-fixing loops burn through their balance, and a Trustpilot record -- 1.5 out of 5 across 202 reviews, 84% one-star -- that is hard to wave away as noise next to a much friendlier G2 number. Both things are true about the same product."
None of that erases what Bolt is good at: going from a chat prompt to a deployed, shareable app faster than almost anything else in this category, with a real free tier and a genuinely novel execution model underneath it. It does mean I'd budget carefully, read the release notes before assuming your plan works the way an older blog post describes, and keep a close eye on the token meter the first time Bolt starts "fixing" something on its own.
One chat box, one in-browser runtime, three agents to choose from.
Bolt.new is StackBlitz's AI app-building product: you describe what you want in a chat box, and Bolt writes, runs, and previews a full-stack app entirely inside a WebContainer running in your own browser tab -- no server-side sandbox spin-up, no waiting on a remote VM. As of this check the agent picker offers three options (Forge, Standard, Max), Bolt Cloud provides built-in databases/auth/storage/hosting, and MCP-based Connectors let Bolt read and write to tools like Notion, Linear, and GitHub.
The one thing worth flagging up front: "Bolt" and "StackBlitz" are often used interchangeably in coverage, but they're not fully merged. stackblitz.com is still a live, separate product (the original cloud IDE); Bolt.new's own Terms of Service and Privacy Policy links still point to stackblitz.com domains as of this check. Bolt has clearly become the company's much bigger, more prominent brand -- it just hasn't legally absorbed the older product.
One of the fastest revenue ramps on record -- and an open-source repo the company itself walked away from.
$105.5M round, Jan 23, 2025, led by Emergence Capital & GV.
As of Dec 2025; no confirmed Series C as of this check.
Company-reported, Mar 2025 -- third-party corroborated, not audited.
May 2025 (company) → Dec 2025 (third-party estimate).
| Date | Event | Result |
|---|---|---|
| 2017 | StackBlitz founded (Eric Simons & Albert Pai, San Francisco) | Browser-based IDE; WebContainers R&D begins |
| Oct 3, 2024 | Bolt.new launches with a single tweet | ~$1M ARR in week one (company-reported) |
| Nov 2024 - Mar 2025 | Revenue ramp | $4M ARR (30 days) → $20M ARR (60 days) → ~$40M ARR (~5 months) |
| Jan 23, 2025 | Series B, $105.5M, led by Emergence Capital & GV | Valuation $700M |
| Sep 2025 - Oct 2025 | Bolt Cloud + Bolt v2 launch | Claude Agent becomes default LLM; built-in DB/auth/hosting |
| Dec 9, 2025 | Listed on AWS Marketplace | Enterprise procurement route opens |
| May 5, 2026 | Microsoft Azure/365 partnership announced | Marketplace listing, BYOC to Azure, Teams integration |
| Sep 14, 2026 | Bolt Forge launches | Third agent, open-weight models, separate token pool |
stackblitz/bolt.new, shows 16,568 stars and 14,818 forks (GitHub, live) -- but it was last pushed Dec 17, 2024 and now carries 8,187 open issues: StackBlitz effectively stopped maintaining its own seed repo about three months after publishing it. Active open-source development continued instead under the community-run bolt.diy fork (formerly oTToDev), which StackBlitz later adopted into its own stackblitz-labs GitHub org: 19,918 stars, 10,586 forks, 143 open issues, last pushed Feb 7, 2026. Stars/forks measure awareness of two different codebases, not paying Bolt.new customers -- worth keeping separate.G2 likes the speed. Trustpilot tells a much rougher story about billing and support.
Reviewers consistently praise the intuitive chat-to-app flow and fast prototyping, plus smooth integration with Supabase and Vercel/Netlify -- while noting the product "struggles with more complex projects, requiring additional prompts and manual adjustments" and flagging dependency issues that complicate getting generated code into an existing repo.
1.5 out of 5 stars, with 84% of reviews at one star. The dominant complaint pattern: tokens consumed fixing bugs Bolt itself introduced (one Product Hunt reviewer estimated roughly half of all tokens spent went to Bolt's own errors), and customer support described as "limited to bot responses."
A founder's account: an agent run burned roughly 10M tokens making destructive changes, plus an undisclosed Netlify deployment that silently broke a Vercel payment integration; a support reply arrived six days later with generic language, and legal/LinkedIn outreach to executives went unanswered. One user's account, not independently arbitrated -- but a real, citable, negative data point.
The sticker price is $25. The real number depends on how many mistakes get fixed on your dime.
Free
300K tokens/day, 1M/month.
- Public & private projects, unlimited databases
- Website hosting, up to 333K web requests
- Bolt branding on published sites
Pro
Starts at 10M tokens/mo, no daily cap.
- Unused tokens roll over to next month
- No Bolt branding, custom domains, SEO boost
- Higher token bundles & top-ups exist above $25
Teams
Everything in Pro, plus admin.
- Centralized billing, access management
- Private NPM registries, per-package prompts
- Org-wide project sharing
Enterprise
Advanced security, compliance, 24/7 support.
- SSO, audit logs, dedicated account manager
- BYOC: your own AWS or Azure tenant
- HIPAA/FedRAMP-ready deployment options
Exact wording pulled from the live pricing page, Sep 28, 2026. Third-party reviewers (Taskade, Banani) report that step-up token bundles inside Pro and one-off token top-ups can push a heavy month well past $25 -- toward $50-$100 or more for large projects; JAVIS did not independently verify those higher-tier figures against a live checkout.
Your answer depends on how much you trust a chat box to self-correct.
Free tier -- 300K tokens/day is enough for a small prototype; watch the daily reset.
Pro, $25/mo, 10M tokens with rollover -- set a personal token budget alarm before you start.
Try Bolt Forge inside Pro -- open-weight models, a separate token pool, promotional 50x usage through Oct 14, 2026.
Teams, $30/user/month -- centralized billing plus org-wide sharing.
Enterprise, custom -- the tier the Microsoft and AWS Marketplace deals target.
Bolt isn't the right primary tool -- WebContainers runs JS/Node.js; Python support is experimental and pip-less.
The thing that actually differentiates Bolt: your browser is the server.
WebContainers is StackBlitz's own invention -- by its own account, four-plus years of engineering to build "a WebAssembly-based operating system powerful enough to run Node.js, entirely inside your browser." In practice that means when Bolt "runs" your app, npm installs, dev servers, and the live preview all execute client-side, in your browser's own security sandbox, with no server-side code-execution step in between. On top of that runtime sits a genuinely new decision as of this check: which of three agents writes your code. Standard (Claude Sonnet 4.6, the default since April 2026) balances speed, cost, and reasoning depth. Max (Claude Opus 4.7, the most capable option since mid-April 2026) is for complex, long-running tasks. Bolt Forge, which shipped Sep 14, 2026 -- about two weeks before this review -- routes to open-weight models (GLM 5.3, GLM 5.3 Flash, Kimi K3, and an experimental DeepSeek v4 Pro) from a completely separate token pool, with Pro subscribers getting a promotional "up to 50x more usage at no extra cost" through Oct 14, 2026.
| Standard | Max | Forge | |
|---|---|---|---|
| Model(s) | Claude Sonnet 4.6 | Claude Opus 4.7 | GLM 5.3 / GLM 5.3 Flash / Kimi K3 / DeepSeek v4 Pro (experimental) |
| Best for | Everyday building, default choice | Complex, long-running reasoning tasks | High-volume exploration/drafting |
| Token pool | Shared plan allowance | Shared plan allowance | Separate pool, no daily limit |
| Cost trade-off | Standard rate | Higher effective cost per task | Promo: up to 50x usage through Oct 14, 2026 |
| Data note | Standard privacy terms | Standard privacy terms | Opt-in: de-identified sessions can help train open models (Arcee AI partnership) |
- WebContainers is a real, hard-to-replicate technical moat -- confirmed by StackBlitz's own multi-year build story, not marketing language.
- Official claims of "98% fewer error loops" and "1,000x bigger projects" since Bolt v2 (Oct 2025) are attributable, sourced claims from the launch post, not independently benchmarked by JAVIS.
- Three real agent tiers, including a genuinely cheaper open-weight option, is more model choice than most rivals in this category offer.
- WebContainers means JavaScript/Node.js first -- Python support is explicitly described as experimental and pip-less, so real Python ecosystems (numpy, pandas, most ML stacks) aren't usable inside Bolt.
- Forge's opt-in data-sharing for open-model training is a real trade-off to read closely, even though it's opt-in.
Official Bolt.new "Demo Day" video: builders present what they shipped end-to-end on the platform.
Published by the official "Bolt•new" YouTube channel (youtube.com/@Boltdotnew). A separate candidate video featuring Bolt's COO ("From Nearly Shutting Down to $40M in 5 Months") was checked the same way and rejected: YouTube lists the author as "TwoSetAI", a third-party interview channel, not Bolt's own.
Prompt in, browser-run app out -- the review step is where trust either gets earned or lost.
The self-correcting debug loop is genuinely useful when it works — StackBlitz's own numbers claim a 98% reduction in error loops since Bolt v2. The documented complaint pattern (Reddit, Product Hunt, the HN incident above) is what happens when it doesn't: the same loop keeps consuming tokens fixing its own mistakes, and there's no built-in spend cap that stops it mid-run.
The sandbox is the browser tab you're already looking at.
Because WebContainers execute entirely client-side, there's no separate remote build server to wait on for the initial run -- the trade-off is that everything is bounded by what a browser sandbox can do. Deployment is more flexible: Bolt Hosting publishes to a free bolt.host subdomain, Netlify remains a supported target (and as of May 2026 you can edit a published Netlify URL without a full redeploy), custom domains are available on Pro and up, and Expo integration adds iOS/Android builds with a QR-code preview for mobile testing directly from the same chat interface.
- Instant preview with no cold-start wait for a remote container
- Zero-setup local dev experience -- nothing to install to get started
- Figma-to-code and Expo mobile preview both live inside the same tab
- Non-JS backends (Python, PHP, Go, etc.) aren't first-class -- Python is experimental and pip-less
- Heavy compute or long-running background jobs don't fit a browser sandbox model well
Connectors are Bolt's name for MCP, and they shipped eight defaults out of the gate.
Since the March 11, 2026 "Introducing Connectors" launch, Bolt runs on the Model Context Protocol as its integration layer: any tool with a remote MCP server can become a Connector with no custom integration work. Eight ship by default -- Notion, Linear, GitHub, Miro, Sentry, Context7, Granola, and Jira -- and you can add a custom connector for anything else by supplying its MCP server URL, transport type, and credentials.
There's no separate Bolt API bill -- your real variable cost is the token meter.
| What you're doing | What you pay |
|---|---|
| Evaluating on the free tier | $0 -- 300K tokens/day, 1M/month, resets on schedule |
| Building on Pro | $25/mo base -- starts at 10M tokens/mo, unused tokens roll over |
| Heavy usage on Pro (reported) | Higher token-tier step-ups + one-off top-ups -- third-party sources report real months landing well above $25 for large projects |
| Experimenting via Bolt Forge | Separate pool, promo "up to 50x usage" through Oct 14, 2026 for Pro subscribers |
| Rolling out to a company | Teams ($30/user/mo) or Enterprise (custom), plus admin/security controls |
Real certifications, a genuinely sandboxed runtime, and one opt-in trade-off worth reading closely.
| Control | What's actually in place |
|---|---|
| Certifications | SOC 2 Type 2 certified; GDPR and CCPA compliant (per bolt.new/platform/security) |
| Encryption | AES-256 or better at rest; TLS 1.2+ over HTTPS in transit |
| Execution isolation | Every project runs in its own browser-level WebContainer sandbox, separated from other projects and Bolt's own infrastructure |
| Model training | Enterprise/Team customers under an MSA or DPA are excluded from model training and data sharing |
| Enterprise deployment | BYOC -- run Bolt inside your own AWS or Azure tenant, full infrastructure isolation, no shared compute |
| Monitoring | Automated scanning/intrusion detection around the clock; independent third-party assessors test annually |
It's worth separating two different kinds of "trust" here: the security architecture (sandboxing, encryption, certifications) is genuinely solid and independently verifiable against the official page. The support/billing-dispute pattern documented in the reviews section is a different, softer kind of risk -- and it's the one I'd weight more heavily before committing real budget.
Best for people who want speed and can absorb a surprise, not people who need budget certainty.
| Situation | Why Bolt fits | Likely plan |
|---|---|---|
| Non-technical founder validating an idea | Chat-to-running-app, no local setup, genuinely fast first result | Free, then Pro |
| Agency building client prototypes | Figma import + fast iteration + custom domains | Pro / Teams |
| Enterprise standardizing on one AI builder | Microsoft/AWS Marketplace procurement, BYOC, SSO | Enterprise, custom |
| Budget-sensitive experimenter | Bolt Forge's separate pool + promo usage through Oct 14, 2026 | Pro (using Forge) |
| Developer needing Python/non-JS backends or hard cost predictability | Not Bolt's strength -- see Limitations and Alternatives | Consider Replit or a traditional stack |
Four real gaps, not manufactured ones.
Independent reports across Reddit, Product Hunt, and Hacker News converge on the same complaint: the autonomous debug loop can consume a large share of your token budget fixing errors it introduced itself, with no built-in spend cap.
1.5/5 across 202 reviews, 84% one-star, dominated by billing and "bot-only support" complaints -- a much rougher picture than G2's friendlier number, and a gap wide enough to take seriously rather than dismiss as one platform's bias.
WebContainers is the whole reason Bolt is fast, but it also means Python support is explicitly experimental and pip-less -- the real Python data/ML ecosystem isn't usable inside Bolt.
stackblitz/bolt.new (the code StackBlitz itself open-sourced at launch) hasn't been pushed to since Dec 2024 and carries 8,187 open issues; if you want an open, actively maintained version, you need the community-run bolt.diy fork, not the original repo.
Bolt isn't the only chat-to-app builder, and it isn't optimized for everyone's tradeoff.
| If you mostly need... | Compare Bolt with... |
|---|---|
| The most full-stack, non-technical-friendly builder with auth/DB/payments wired by default | Lovable -- read our review |
| The cleanest frontend code output and one-click Vercel deploys, no backend needed | v0 -- read our review |
| A general-purpose cloud IDE/agent across many languages, not just Node.js | Replit |
| A professional developer-first editor with deep existing-codebase context, not green-field prototyping | Cursor -- see our Cursor review |
| The open-source, run-anywhere, bring-your-own-LLM version of Bolt's original UI | bolt.diy (community fork, MIT, actively maintained) |
No affiliate relationship shapes this review.
JAVIS has not established a publisher affiliate program with StackBlitz/Bolt.new. Every CTA in this article points to the official product.
Go to the official page.
Open Bolt.newRead the alternative that matches your real question.
Choose the next articleBolt makes the most sense next to the other chat-to-app builders and the Claude models underneath it.
Bolt, Lovable, and v0 are worth reading together because they're the three tools most people actually shortlist against each other. Here's what I'd read next.
Questions people are actually searching right now
Is Bolt.new free to use?
Yes. The Free plan gives you 300K tokens/day and 1M tokens/month, unlimited databases, and website hosting, with Bolt branding on published sites. No credit card required to start.
What is Bolt Forge?
A third agent option, launched Sep 14, 2026, that routes to open-weight models (GLM 5.3, GLM 5.3 Flash, Kimi K3, and an experimental DeepSeek v4 Pro) instead of Claude. It draws from a separate token pool with no daily limit, and Pro subscribers get a promotional "up to 50x more usage" through Oct 14, 2026. Opting in lets de-identified sessions help train open models via StackBlitz's partnership with Arcee AI.
Does Bolt.new support Python or PHP?
Only in a limited, experimental way. WebContainers -- the WebAssembly-based runtime Bolt is built on -- natively executes JavaScript/Node.js. Bolt's Python support is described as experimental and pip-less, so real Python ecosystems (numpy, pandas, most ML libraries) aren't usable inside the sandbox.
Is Bolt.new the same company as StackBlitz?
Bolt.new is StackBlitz's product, built on its WebContainers technology. As of this check, stackblitz.com is still a separate, live product (the original cloud IDE), and Bolt.new's own Terms of Service/Privacy Policy links still point to stackblitz.com domains -- so it's more accurate to say Bolt became the company's dominant consumer brand than that StackBlitz fully renamed itself.
What happened to Bolt's open-source version?
StackBlitz open-sourced the original code at launch (stackblitz/bolt.new, MIT license) but hasn't pushed to that repo since Dec 17, 2024, and it now has 8,187 open issues. Active open development continued instead through the community-run bolt.diy fork (formerly oTToDev), which StackBlitz later folded into its own stackblitz-labs GitHub org -- that fork is the actively maintained option if you want to run an open version yourself.
Is Bolt.new secure enough for enterprise use?
The stated architecture is solid: SOC 2 Type 2, GDPR/CCPA compliance, AES-256 at rest, TLS 1.2+ in transit, per-project browser sandboxing, and BYOC deployment into your own AWS or Azure tenant for Enterprise customers. Separately, the real-world support/billing-dispute pattern documented in user reviews is a different kind of risk worth weighing on its own.
How does Bolt.new compare to Lovable and v0?
Independent 2026 comparisons consistently describe Lovable as the most full-stack-by-default (Supabase, auth, Stripe wired in automatically, more design-polished output), v0 as the most Vercel-native with the cleanest frontend code but no backend story, and Bolt as the most framework-flexible with a genuinely novel in-browser execution model -- at the cost of more variable code quality and the token-burn pattern documented above.
Did Bolt.new partner with Microsoft?
Yes -- announced May 5, 2026. Bolt.new is listed on Microsoft Marketplace, supports deployment into a customer's own Azure tenant, integrates with Microsoft 365 and Teams for collaborative prompting, and supports Microsoft Entra and Defender for enterprise security. It's also listed on AWS Marketplace as of Dec 9, 2025.
Where this came from, and when it was checked.
Pricing (Free/Pro/Teams/Enterprise): bolt.new/pricing, Sep 28, 2026.
Product homepage claims, security architecture: bolt.new and bolt.new/platform/security, official, checked Sep 28, 2026.
Bolt v2 (98% fewer error loops, 1,000x bigger projects, Claude Agent default): bolt.new/blog/bolt-v2, official announcement, Oct 2, 2025.
Bolt Forge (open-weight agent, Arcee AI partnership): bolt.new/blog/what-is-bolt-forge, official, Sep 14, 2026; corroborated by support.bolt.new/release-notes.
Connectors/MCP: bolt.new/blog/introducing-connectors, official, Mar 11, 2026.
Microsoft partnership: bolt.new/blog/news-press-release-bolt.new-on-microsoft-azure, official, May 5, 2026, including direct quotes from Microsoft's Mike Hulme and Bolt CEO Eric Simons.
AWS Marketplace listing: Benzinga and EZ Newswire coverage, Dec 9, 2025; corroborated on aws.amazon.com/marketplace.
GitHub stats (both repos): GitHub data for github.com/stackblitz/bolt.new and api.github.com/repos/stackblitz-labs/bolt.diy, Sep 28, 2026.
StackBlitz company history (2017 founding, WebContainers R&D): Contrary Research company report, Business Insider/Facebook coverage, cross-checked, Sep 2026.
Funding/ARR/user milestones: Sacra, growthunhinged.com, getpanto.ai -- third-party/company-disclosed, not independently audited by JAVIS, flagged as such throughout.
G2 rating/review pattern: search-result aggregation of the live G2 page (the page could not be read directly), checked Sep 2026.
Trustpilot rating: secondary source (superdesign.dev) reporting a direct read of the live Trustpilot page, checked ~Sep 1, 2026 per that source.
Named user incident: news.ycombinator.com/item?id=46321594, Sep 2026 -- one user's account, presented as such.
Python/backend language limitation: support.bolt.new/concepts/supported-technologies and independent technical write-ups, cross-checked against StackBlitz's own WebContainers description.
Official video: youtube.com/watch?v=8jbIMr6yRMM, authorship confirmed on YouTube (author "Bolt•new", channel @Boltdotnew); one other candidate video was checked the same way and rejected as non-official.
