Can Claude Opus 4.8 Build a Full SaaS App Alone?
We built a real SaaS app with Claude Opus 4.8 and tracked every phase. What it nailed, where it broke, the time and cost, and the honest verdict.
We built a real SaaS app with Claude Opus 4.8 and tracked every phase. What it nailed, where it broke, the time and cost, and the honest verdict.
Claude Opus 4.8 can build roughly 85% of a production SaaS app on its own — auth, database, CRUD, billing wiring, and a polished UI — but the remaining 15% is the part that decides whether the app survives real users, and that part still needs a human in the loop. "Alone" is the wrong frame. "With a competent operator and almost no hand-written code" is the accurate one.
To test the claim properly we built a real, deployable SaaS product end to end and logged every phase: what Claude produced unassisted, where it stalled, how long it took, what it cost, and what a security pass surfaced. This is the full account, with the numbers.
The summary first, so you know where this lands before the detail:
| Capability | Grade | Notes |
|---|---|---|
| Scaffolding & project setup | A | Near-flawless. Next.js, TypeScript, Tailwind in minutes. |
| UI & frontend | A | Polished, responsive, consistent. The strongest area. |
| CRUD & business logic | A- | Reliable for standard patterns; needs review on edge cases. |
| Database schema & queries | B+ | Sensible schemas; occasional missing indexes and N+1 queries. |
| Authentication | B | Works, but defaults were not production-hardened. |
| Payments / billing | B- | Wired Stripe correctly; missed webhook idempotency until prompted. |
| Security | C+ | The weakest area. Several issues a human had to catch. |
| Testing | B | Wrote good tests when asked; did not volunteer coverage. |
| Deployment & DevOps | C | Could explain steps; execution needed human hands and accounts. |
| Architecture decisions | B | Good defaults, but will not push back on a bad spec unprompted. |
Frontend and scaffolding are effectively solved. Security and deployment are where the human earns their keep. The middle — logic, data, auth, billing — is excellent with review and risky without it.
To keep the test honest, we defined the product before touching Claude and did not move the goalposts mid-build.
| Parameter | Choice |
|---|---|
| Product | A team feedback-board SaaS (think a lightweight Canny clone) |
| Core features | Auth, organizations, boards, posts, voting, comments, admin, Stripe billing |
| Stack | Next.js 15, TypeScript, Tailwind, Postgres (Supabase), Stripe |
| Tooling | Claude Code (CLI agent) + Claude.ai for design Artifacts |
| Human role | Operator only — prompts, reviews, approvals. No hand-written feature code. |
| Definition of done | Deployed, a real user can sign up, create a board, and pay for a plan |
The "operator only" constraint is the important one. The human could prompt, review, reject, and run commands, but could not write feature code by hand. That is the closest fair test of "alone."
Nine phases, start to deployed. The intervention column is where the "alone" claim gets tested.
| Phase | What Claude did unassisted | Human intervention required | Time |
|---|---|---|---|
| 1. Scaffold | Full Next.js + TS + Tailwind project, folder structure, config | None | 10 min |
| 2. Database schema | Designed all tables, relations, RLS policies | Added two missing indexes after review | 35 min |
| 3. Auth | Email + OAuth via Supabase Auth, protected routes | Hardened session config, fixed a redirect gap | 50 min |
| 4. Core UI | All pages, components, responsive, empty/loading states | Minor copy and spacing tweaks | 70 min |
| 5. CRUD & logic | Boards, posts, votes, comments with optimistic updates | Fixed a vote race condition flagged in review | 90 min |
| 6. Admin | Role-gated admin panel, moderation actions | Tightened permission checks | 45 min |
| 7. Billing | Stripe checkout, plans, customer portal | Added webhook idempotency + signature verification | 75 min |
| 8. Tests | Unit + integration tests for core flows (on request) | Asked for coverage it had skipped | 60 min |
| 9. Deploy | Wrote deploy config + step-by-step instructions | Created accounts, set secrets, ran the deploy | 40 min |
Total wall-clock: about 7.5 hours across two sittings, producing a deployed, working app. Every line of feature code came from Claude. Every architectural and security judgment call had a human behind it.
The failures were predictable and clustered. None were dealbreakers, but every one would have shipped if a human had not been reviewing.
| Issue | Severity | Would it have shipped? |
|---|---|---|
| Stripe webhooks lacked idempotency | High | Yes — duplicate charges under retries |
| One API route missing an auth check | Critical | Yes — data exposure |
| Vote endpoint race condition | Medium | Yes — inconsistent counts |
| Missing DB indexes on hot queries | Medium | Yes — slow at scale, invisible early |
| No rate limiting anywhere | High | Yes — abuse vector |
| Secrets logged in a debug statement | High | Maybe — caught in review |
| Did not volunteer tests | Low | N/A — needed prompting |
The pattern is clear: Claude builds what you ask for, correctly, but does not independently think like an adversary or a site-reliability engineer. It does not ask "what happens under load, retries, or attack?" unless you make it. That gap is the whole reason "alone" fails.
We ran a dedicated security review as its own phase — and it was the most revealing part of the experiment. Prompted explicitly to audit its own code as an attacker, Claude found most of its own issues. Left to build features without that prompt, it shipped them.
| Security check | Default build | After explicit audit prompt |
|---|---|---|
| Auth on every protected route | 1 route missed | Fixed |
| Input validation / sanitization | Partial | Comprehensive |
| Rate limiting | None | Added on auth + write routes |
| Webhook signature verification | Missing | Added |
| Secrets handling | One leak in logs | Removed |
| SQL injection surface | Safe (parameterized) | Confirmed safe |
| RLS / row-level access | Mostly correct | Tightened 2 policies |
The lesson is actionable: security is a prompt you must run, not a property you get for free. Build the app, then make Claude audit it as an adversary before anything touches real users.
How the 7.5-hour Claude-assisted build compares to the alternatives for the same scope:
| Approach | Time to deployed MVP | Cost |
|---|---|---|
| Claude Opus 4.8 + operator | ~7.5 hours | ~$30 in API + $20 plan |
| Solo senior developer | ~3–5 days | ~$2,000–$4,000 in time |
| Small agency / freelance team | ~2–4 weeks | ~$8,000–$25,000 |
| No-code platform | ~2–3 days | Lower upfront, capped ceiling |
API token usage for the full build landed around $28, plus the operator's Claude plan. On any cost-per-outcome basis, the AI-assisted route is roughly an order of magnitude cheaper and faster than the alternatives for an MVP of this scope.
| Cost component | Amount |
|---|---|
| Claude API tokens (build) | ~$28 |
| Claude plan (operator) | $20–$200 / month |
| Supabase (free tier) | $0 to start |
| Hosting (Vercel hobby/pro) | $0–$20 / month |
| Stripe | Per-transaction only |
The build quality tracks directly to the spec quality. This is the opening prompt we gave Claude Code — note how much it constrains up front, which is what kept the 7.5-hour run coherent.
You are a senior full-stack engineer building a production SaaS app. Product: a team feedback-board SaaS (lightweight Canny clone). Stack: Next.js 15 (App Router), TypeScript, Tailwind, shadcn/ui, Supabase (Postgres + Auth), Stripe. Build plan — propose it first, then execute phase by phase, pausing for my approval after each: 1. Scaffold the project with a clean structure. 2. Design the full database schema with relations and row-level security. 3. Implement auth (email + OAuth) with protected routes. 4. Build the UI for every page with real copy and all interactive states. 5. Implement CRUD + business logic (boards, posts, voting, comments). 6. Add a role-gated admin panel. 7. Wire Stripe billing (checkout, plans, customer portal, webhooks). 8. Write unit + integration tests for the core flows. 9. Produce deployment config and step-by-step deploy instructions. Rules: use the 8px spacing scale, validate all inputs, never log secrets, parameterize all queries, and flag any security or scaling concern as you go. After each phase, summarize what you built and what you deliberately deferred.
Two details made the difference: forcing a phase-by-phase plan with approval gates, and the standing rule to flag security and scaling concerns. The second turned Claude into a partial reviewer of its own work. For more patterns like this, see our Next.js prompts for Claude Opus.
The precise answer, by who is asking:
| If you are a... | Can Claude 4.8 build your SaaS? |
|---|---|
| Experienced developer | Yes — it is a 10x force multiplier. You review and direct; it writes. |
| Technical founder (can read code) | Mostly — you can ship an MVP, but get a security review before launch. |
| Non-technical founder | To a working prototype, yes. To a safe production launch, not alone. |
| Designer / PM | A convincing, clickable prototype — yes. Production — pair with a dev. |
"Alone" is a marketing word. The real model is Claude as the builder, a human as the architect and security reviewer. In that pairing it produced a deployable SaaS in 7.5 hours for under $30 — work that costs thousands and weeks otherwise. The 15% it cannot be trusted to own alone is exactly the 15% that protects your users, and pretending otherwise is how AI-built apps leak data.
The honest framing: Claude Opus 4.8 does not replace the engineer. It replaces the typing. The judgment still has to be human — and for now, that is the right division of labor.
Browse the full PromptsRush blog, our prompt library, and the AI model directory.
10 questions answered
How Developers Are Using Claude 4.8 for Vibe Coding
May 30 · 10 min
How to Use Claude Design: 25+ Working Prompts (2026)
May 28 · 14 min
Claude 4.8 vs Claude 4.7: What Actually Improved (2026 Benchmarks)
May 28 · 11 min
10 Things Claude Opus 4.8 Can Do Better Than GPT-5.5
May 28 · 8 min
Claude Opus 4.8 vs GPT-5.5: Which AI Model Wins in 2026?
May 28 · 9 min