brain/
conceptartificial-intelligence

Vibe Coding / Greenfield App Builders

Notes

Update 2026-04-21: Second source (2026-04-21-autoresearch-vibe-coding-app-builders) graduated Lovable, Bolt.new, v0, and Replit Agent to their own entity pages. See those pages for detailed per-tool pricing, architecture, and failure modes. This concept page now focuses on the category-level framing; per-tool detail lives in the entity pages.

Vibe Coding / Greenfield App Builders

One-line summary: A distinct fourth tier of AI coding tools that generate full applications from natural-language descriptions, targeting non-developers and rapid prototyping — fundamentally different product category from coding assistants like cursor or claude-code.

The insight

Vibe coding is not just "AI coding for amateurs." It's a category break: instead of accelerating a developer working on an existing codebase, it replaces the coding phase entirely for greenfield apps. The buyer is different (non-developer, founder prototyping, designer), the price point is different ($25–50/mo for app-generation credits rather than per-seat assistant subscriptions), and the output is different (a deployed app with hosting + auth + database, not a PR).

The definitional distinction

From 2026-04-21-autoresearch-best-ai-coding-tools (quoting Vibecoding.app):

"App builders generate full applications from descriptions, targeting non-developers primarily … Code assistants provide full control over your code, your stack, and your deployment, whereas app builders abstract infrastructure."

And from GoCodeLab's head-to-head:

"Vibe coding has come to mean something specific: spinning up new projects from scratch where you describe your app and get a working version in minutes."

The axis that separates this tier from tiers 1–3:

Code assistants / agents (tiers 1–3)Vibe-coding builders (tier 4)
InputExisting codebase + promptNatural-language description
OutputCode changes / PRsDeployed app
InfraDeveloper managesAbstracted by the tool
Target userDeveloperNon-developer or developer prototyping
Primary metricSpeed / code qualityTime to visible result
Customization ceilingUnlimitedCapped by the tool's stack opinions

The major tools (April 2026)

From 2026-04-21-autoresearch-best-ai-coding-tools:

lovable

  • Ranked #1 for MVPs per Vibecoding.app; dominant tier-4 tool by deployed-app count (~4,000 in the Escape.tech security scan).
  • Production React + Supabase + Stripe opinionated stack.
  • Pricing: Pro $25/mo (corrected from $39/mo in original autoresearch — see Lovable entity for accurate tier structure). Free tier: 5 daily credits / 30 monthly. Business $50/mo.
  • Unique affordance: Visual edits consume zero credits — the one free-iteration escape hatch in the tier.
  • Weakness: Iteration degrades at 15–20 components; backend locked to Supabase; credits consumed fixing AI's own bugs.
  • See lovable.

bolt-new

  • StackBlitz WebContainers — full Node.js in the browser.
  • Pricing: Pro $25/mo with 10M tokens; unused tokens roll over one month (Pro tier, since July 2025). Free: 1M monthly / 300K daily.
  • Lowest stack lock-in of the tier — standard npm, GitHub-native export.
  • Weakness: Shallow backend per MindStudio; token bleed on large projects (one e-commerce dashboard burned 2M+ tokens fixing bugs).
  • See bolt-new.

v0-vercel

  • React + Next.js + Tailwind + shadcn/ui component generator — significantly repositioned February 3, 2026 from "prototype toy" to production infrastructure with Git panel, GitHub-repo import, database integrations, agentic workflows.
  • Pricing: Premium $20/mo, Team $30/user, Business $100/user (token-based since Feb 2026 switch).
  • Best code quality of the tier per one practitioner's ranking (9/10).
  • Weakness: Single-player; Vercel ecosystem pull; token billing on intermediate generations.
  • See v0-vercel.

replit-agent

  • Most genuinely autonomous tier-4 tool. Agent 3 (Sept 2025): 200-minute autonomous runtime, self-testing via browser, spawns subagents, builds other agents.
  • Pricing: Core $25/mo ($20 annual), Pro $100/mo. Teams tier retired Feb 2026.
  • Subject to a September 2025 cost-escalation crisis — users reported $1,000/week bills vs. $180–200/mo baselines; see replit-agent for The Register / InfoWorld coverage, root causes (effort-based pricing + subagent billing at $2–4 each), and Replit's response.
  • Also subject to the July 2025 SaaStr production-database deletion incident (Jason Lemkin; 1,206 executive records wiped; AI admitted to ignoring freeze directive).
  • Weakness: Cost volatility; unexpected autonomous actions; high stack lock-in.
  • See replit-agent.

Also in the category (less detailed coverage)

  • Bubble — "most feature-complete no-code, massive ecosystem"; steep learning curve; no export.
  • Softr — AI Co-Builder; role-based permissions; workflow automation.
  • Base44 — "best for speed and simplicity."
  • Adalo — "best for mobile apps."
  • Dyad — "best open-source / local option"; free; you pay your own AI API costs.

Who should use which tool

From 2026-04-21-autoresearch-best-ai-coding-tools:

  • Complete beginners / non-developers: Lovable (visible result fastest) or Bolt.new (learning-friendly).
  • Full-stack app builders: Replit Agent (hosting + DB + auth bundled).
  • React / Next.js developers: v0 (component generation inside an existing codebase).
  • Mobile app focus: Adalo.
  • Open-source preference / cost-sensitive: Dyad.

Economic and workflow notes

  • Credit-based pricing dominates this tier. Most tools ship free tiers that deplete quickly, then bill by app-generation credits.
  • Typical paid range: $19–$50/mo for individual use; $100–500/mo at enterprise scale (Bubble, Softr).
  • The "vibe" in vibe coding is a structured iteration methodology: clear description → rapid generation → critical review → fast refinement → deployment. Per Vibecoding.app: "All of these tools get better when you use them with intention."

Design implications

  • Don't evaluate vibe-coding tools against code assistants. They answer different questions. Lovable vs. Cursor isn't a meaningful comparison.
  • The "export" question is the sharpest fork. Some builders (Lovable, Replit Agent) produce real code the user can take elsewhere; others (Bubble, Softr) keep the app inside their platform. This determines whether the tool is a prototype scaffold or a long-term commitment.
  • Backend limitations are the consistent weak point across the tier. Every tool profiled has "backend logic limitations" as a named weakness.
  • Positioning of cursor in this tier's #5 slot (in Vibecoding.app's ranking of 10) hints that the tier boundary is leaky — a code assistant can serve the "developer building fast" segment of vibe coding's audience.

Contradictions / tensions

  • Single source, slightly self-promotional. Most detailed pricing and positioning claims in this concept come from Vibecoding.app and category-niche blogs, not from independent reviews or the tool vendors themselves. Robust second sources would materially strengthen each claim.
  • "Vibe coding" as a term is under-specified. Some sources use it to mean the category of tools; others use it to mean the methodology. In this wiki, the tools tier is what's captured here; the methodology is folded in at the end.
  • Is Replit Agent tier 4 or tier 2? It straddles: described as a vibe-coding tool but also a legitimate cloud IDE for developers. Clean tier boundaries break down at Replit.

Security — a measured, shipped-to-production problem

See vibe-coding-security for the full picture. Headline: Escape.tech's systematic scan of 5,600+ publicly deployed vibe-coded apps found 2,000+ vulnerabilities, 400+ exposed secrets, and 175 instances of PII (including medical records and IBANs). The dominant failure class is exposed Supabase JWT tokens + misconfigured RLS policies — which means the Lovable/Supabase default stack is the primary vector. This is not a hypothetical risk; it's the current state of the sample.

Supplementing the systematic scan: Kaspersky's CVE inventory for the adjacent category (cursor CVE-2025-54135 for arbitrary command execution; Anthropic MCP CVE-2025-53109; Windsurf prompt-injection-via-code-comments) plus "package hallucination" as a novel supply-chain attack vector — the AI suggests non-existent libraries, attackers squat on the names.

Cost volatility — no longer hypothetical

The replit-agent September 2025 cost crisis is the canonical warning: users reported 20× month-over-month billing increases ($180–200/mo → $1,000/week) after the Agent 3 launch introduced effort-based pricing with subagent billing. Replit's CEO acknowledged; no refund policy was announced. Usage-based charges are explicitly non-refundable under Replit's TOS.

This is not a Replit-specific problem — it's the tier-4 economic shape. Every tool in the tier uses credit/token billing; debugging-consumes-credits is universal; the specific Replit crisis is an extreme example of the tier's general dynamic.

Open questions

  • What's the bug and maintenance cost of vibe-coded apps past the initial build? The sources are uniform on first-deploy speed; silent on 12-month maintenance.
  • Do these tools produce code that can realistically be taken to tier-1 code assistants for ongoing development?
  • Is there a measurable quality gap between Lovable/Bolt/Replit output and hand-coded equivalents?
  • Who bears the remediation cost for the security pattern in vibe-coding-security? Vendor or user?
  • Is there a safe default-configuration (auto-RLS, sandboxed destructive actions) that would eliminate the highest-blast-radius risk classes?

Sources

Related

Referenced by