“Vibe coding” — a term Andrej Karpathy popularized in February 2025 to describe building software by describing what you want in natural language and accepting the AI’s output without closely reading the code — has gone from a tongue-in-cheek tweet to a dictionary entry (Merriam-Webster added it in 2025) to a genuine product category. Replit Agent, Lovable, and Bolt.new (built by StackBlitz) are the three platforms most associated with the term: each takes a prompt and produces a working, deployed full-stack app rather than a code suggestion inside an existing editor. That’s a meaningfully different product than an AI coding assistant like GitHub Copilot or Cursor — those help a developer write code faster inside code they already understand; these three are built so a prompt can skip that step entirely. We checked each vendor’s current pricing page and the public security research around this category before writing this piece.
At a Glance
| Platform | What It Produces | Entry Paid Price | Free Tier | Underlying Stack |
|---|---|---|---|---|
| Replit Agent | Full-stack apps built and run inside Replit’s own cloud workspace, with built-in hosting and a database | Core: $20/mo ($18/mo annual) | Limited free usage on Replit’s free plan; Core’s $20 includes model credit | Replit’s own runtime/hosting, Agent orchestrates multiple models |
| Lovable | Full-stack web apps wired to a Supabase backend (auth, database, storage) by default | Pro: from $25/mo (100 monthly credits) | 5 daily credits, capped around 30/month | Supabase for backend/auth/database; React-based frontend |
| Bolt.new | Full-stack apps built and run in-browser via WebContainers, with one-click deploy | Pro: $25/mo | 1,000,000 tokens/mo, 300,000 daily cap, Bolt branding | StackBlitz’s WebContainers (runs Node.js in-browser); deploys to Netlify/other hosts |
Short Verdict
All three platforms do what they advertise: a plain-language prompt becomes a running, deployed application faster than almost any traditional workflow. Replit Agent is the most “full environment” of the three, since Replit already provided cloud dev environments before adding an agent on top. Lovable is the most backend-opinionated, wiring every app to Supabase by default, which is also the source of its most serious documented security incident. Bolt.new is the fastest to a deployed URL because it runs entirely in-browser via WebContainers rather than provisioning a remote environment. None of the three should be treated as producing production-ready code without a security review — the evidence on that point, covered below, is specific and sourced, not just vibes-based skepticism.
What Each Platform Actually Produces
Replit Agent operates inside Replit’s existing cloud workspace product, so an app it builds is already running in a real, persistent environment with its own storage and networking — there’s no separate “deploy” step conceptually distinct from where the code already lives. Agent usage draws on a shared pool of model credit included in each paid plan ($20 worth on Core, $100 on Pro), and Replit’s own pricing page explicitly notes that “Agent is powered by large language models and may occasionally make mistakes” — a disclosure most AI-assistant vendors bury deeper in their terms.
Lovable generates a React-based frontend wired to a Supabase project by default, which means every Lovable app inherits Supabase’s authentication, Postgres database, and file storage out of the box rather than requiring the user to configure a backend separately. That default wiring is also exactly where Lovable’s most significant public security incident originated (see below). Pricing runs on a credit system: each AI action (a UI tweak, a new feature, an auth flow) consumes a fractional credit, with Lovable’s own examples ranging from 0.50 credits for “make the button gray” to 1.20 credits for “add authentication with sign up and login.”
Bolt.new, built by StackBlitz, is the most architecturally distinct of the three: it runs the generated app’s Node.js runtime directly inside the browser using StackBlitz’s WebContainers technology, rather than provisioning a remote VM or container. That’s what lets it go from prompt to a working, previewable app almost instantly, and why its usage is metered in “tokens” (covering reading your prompt and project files, planning, and writing code) rather than a flat per-message credit. Unused paid-plan tokens roll over for up to two months while a subscription stays active; free-plan tokens do not roll over.
Pricing Breakdown
| Platform / Tier | Price | Notes |
|---|---|---|
| Replit Core | $20/mo, $18/mo billed annually | Up to 30 hours chat / 60 projects on Free Mode, plus $20 toward more powerful models |
| Replit Pro | $100/mo, $90/mo billed annually | More Free Mode usage, $100 toward top-tier models plus $100 in additional credits, up to 15 collaborators, 28-day database rollback |
| Replit Enterprise | Custom | SSO, SCIM, audit logs, single-tenant/private networking, dedicated support |
| Lovable Free | $0 | 5 daily credits, capped around 30/month total |
| Lovable Pro | From $25/mo (100 credits) scaling to roughly $2,250/mo (10,000 credits) | Credit packs scale linearly; higher packs trade a worse per-credit rate for more total usage |
| Lovable Business | From $50/mo (100 credits), roughly double the Pro price at every credit tier | Adds team collaboration features on top of the same credit scale as Pro |
| Lovable Enterprise | Custom | Volume-based credit pricing |
| Bolt.new Free | $0 | 1,000,000 tokens/mo, 300,000 daily limit, Bolt branding on deployed apps |
| Bolt.new Pro | $25/mo billed monthly | 10,000,000+ tokens/mo, no daily limit, no branding, custom domains, up to 100MB uploads |
| Bolt.new Teams | $30/member/mo billed monthly | Centralized billing, access management, private NPM registry support |
| Bolt.new Enterprise | Custom | SSO, audit logs, compliance support, dedicated account manager |
Lovable’s credit pricing is the least linear of the three: its per-credit cost rises as you buy more at the Pro tier in some published breakdowns, which makes a heavy-usage team’s actual monthly bill harder to predict from the sticker price alone than either Replit’s or Bolt’s flatter, token/hour-based plans.
What the “Vibe Coding” Critique Is Actually About
The discourse around vibe coding isn’t just generic “AI code can be wrong” skepticism — there’s specific, documented evidence tied to this exact category of tool. The clearest example is Lovable’s own CVE-2025-48757: because Lovable wires apps to Supabase by default, a significant share of apps generated on the platform shipped without Supabase’s Row Level Security (RLS) enabled on their database tables — meaning an app’s entire dataset could potentially be read or modified by anyone who found the right API endpoint, no authentication required. A scan covering 1,645 Lovable-generated apps found 303 vulnerable endpoints spread across 170 of them, a concrete, reproducible number rather than a vague warning, and the issue was severe enough to receive its own CVE identifier.
Independent of that specific incident, Replit Agent has also been directly tested in controlled security research: Tenzai had five agentic coding tools — including Replit — build three identical applications from identical prompts, and found 69 vulnerabilities across the resulting 15 apps. The most consistent failure pattern wasn’t classic injection flaws (no agent produced an exploitable SQL injection or XSS in that test) but access-control and business-logic gaps: every one of the five agents introduced a server-side request forgery (SSRF) vulnerability in a dedicated SSRF test, and researchers called improper authorization on API endpoints “one of the most common issues” across all five tools. Four of five agents also allowed negative-quantity orders when a prompt didn’t explicitly forbid it — a logic flaw no static scanner reliably catches, because nothing about it is syntactically wrong.
A broader, less rigorously documented data point comes from Apiiro, a code-security vendor, whose research (reported via secondary coverage rather than a fully public methodology) claims AI-assisted developers introduce roughly ten times more security vulnerabilities than non-AI-assisted ones, with the largest single increase in privilege-escalation-related flaws. Apiiro’s own framing — “AI is fixing the typos but creating the timebombs” — is a useful summary of the pattern across all of this research: vibe-coding tools are measurably good at producing code that runs and looks syntactically correct, and measurably worse at the parts of security that require understanding intent rather than syntax, like who should be allowed to call a given endpoint. Treat the specific “10x” figure as a vendor-reported headline rather than an independently replicated statistic, since the underlying sample size and method weren’t published in full.
Pros and Cons
Replit Agent
- Pros: fully integrated cloud workspace means no separate hosting step, generous collaborator limits on Pro, transparent model-credit allocation.
- Cons: tied to Replit’s own runtime rather than a portable standard container; included in the Tenzai study’s finding of widespread access-control gaps across agentic tools.
Lovable
- Pros: fastest path to a working backend (auth, database, storage) because Supabase wiring is automatic; granular per-action credit pricing for light usage.
- Cons: that same default Supabase wiring was the direct cause of a CVE affecting hundreds of real deployed apps; credit pricing scales less predictably than the other two at higher usage.
Bolt.new
- Pros: fastest to a live preview via in-browser WebContainers, flat and predictable token-based pricing, generous free-tier token allowance.
- Cons: free-tier apps carry visible Bolt branding; WebContainers’ in-browser execution model has different resource/scaling characteristics than a traditional server-side deploy, which can matter for heavier backend workloads.
Who It’s a Good Fit For
Choose Replit Agent if you want the generated app to live in a persistent, full-featured cloud workspace rather than exporting to separate hosting. Choose Lovable if you specifically want a Supabase-backed app scaffolded fast — and you commit to auditing Row Level Security on every table before shipping. Choose Bolt.new if speed to a shareable preview matters most and you’re comfortable with an in-browser execution model. All three are strongest for prototypes, internal tools, and MVPs where a human developer will review and harden the output before it handles real user data.
Who Should Look Elsewhere
Teams building anything that handles sensitive user data, payments, or regulated information should not ship any of these three platforms’ output straight to production without an independent security review — the Lovable CVE and the Tenzai study’s access-control findings are specific, documented reasons, not generic caution. Teams that need an AI pair-programmer inside code they already understand and maintain — rather than a tool that generates a whole new app from a prompt — are better served by an AI coding assistant like GitHub Copilot or Cursor than by any of these three “vibe coding” platforms.
Bottom Line
Vibe coding in 2026 genuinely works for what it’s built for: turning a plain-language description into a running application faster than a human typing it out by hand. What it does not reliably do, across all three platforms and backed by named, documented incidents and studies rather than general skepticism, is get access control and business logic right by default. That’s a solvable problem with review and scanning — the same playbook that applies to any AI-generated code — but it’s not solved by the platforms themselves yet, and the Lovable CVE in particular shows what happens when that gap reaches production at scale.
FAQ
Who coined the term “vibe coding”? Andrej Karpathy, in a widely shared February 2025 post describing a style of coding where you “fully give in to the vibes” and accept AI-generated code without close review. The term was added to Merriam-Webster’s dictionary later that year.
Was the Lovable security issue theoretical or did it affect real apps? It affected real, deployed apps. The issue has a published CVE (CVE-2025-48757), and an independent scan found 303 vulnerable endpoints across 170 of 1,645 Lovable-built apps checked, caused by missing Supabase Row Level Security.
Do these tools produce code with classic flaws like SQL injection? Not necessarily — in the Tenzai study of five agentic tools including Replit, none of the generated apps contained an exploitable SQL injection or XSS vulnerability. The failures were concentrated instead in access control and business logic: every agent tested introduced an SSRF flaw, and authorization gaps on API endpoints were described as one of the most common issues.
Is the claim that AI-assisted coding produces “10x more vulnerabilities” solid? It comes from Apiiro, a security vendor, via secondary press coverage rather than a fully published methodology and sample size, so treat it as a vendor-reported figure worth taking seriously rather than an independently verified statistic.
Sources
- Replit — Official Pricing Page
- Lovable — Plans and Credits Documentation
- Bolt.new — Official Pricing Page
- NVD / multiple CVE databases — CVE-2025-48757 (Lovable / Supabase RLS)
- Superblocks — “Lovable Vulnerability Explained: How 170+ Apps Were Exposed”
- SecurityWeek — “Vibe Coding Tested: AI Agents Nail SQLi But Fail Miserably on Security Controls” (Tenzai study)
- getCoAI / Apiiro — “Study Finds AI Coding Creates 10x More Security Vulnerabilities”
- Merriam-Webster — “vibe coding” dictionary entry
Discover more from Software Online Review
Subscribe to get the latest posts sent to your email.
