There are two sorts of vibe coding articles on the web proper now.
The primary variety is written by individuals who constructed a to-do app over a weekend and declared software program engineering useless. The second is written by engineers who reviewed one AI-generated codebase, discovered the uncovered API keys, and declared vibe coding a menace.
Each are describing the identical failure: placing AI-generated code within the unsuitable layer of the stack.
This text is concerning the sample that works. We have been operating it in manufacturing for over a 12 months at Loop Followers throughout two shopper merchandise, a blockchain pockets layer, and a rising pile of inner instruments. It allow us to go from 20 builders to three whereas transport quicker than we ever have. I will cowl the structure, the MCP server layer that ties it collectively, the failure modes we hit, and the playbook if you wish to run it your self.
The thesis in a single line
Vibe code the presentation layer. Hand-build the system of report. Join them with open APIs, and expose these APIs to your AI brokers by MCP servers.
Every little thing else on this article is an elaboration of that sentence.
Fast definitions, then we transfer on
Vibe coding — Andrej Karpathy’s time period from early 2025 — is constructing software program by pure language dialog with an AI, evaluating outputs by conduct moderately than studying each line. Instruments like Lovable, Cursor, Claude Code, Replit, Bolt.new, v0, and Windsurf have turned it from a meme into a technique. The excellence from no-code issues: no-code caps you on the platform’s templates; vibe coding emits actual code you personal, can export, and may prolong.
MCP (Mannequin Context Protocol) is an open commonplace, launched by Anthropic, for connecting AI fashions to exterior programs. An MCP server exposes three primitives to any related AI shopper:
- Instruments — actions the mannequin can invoke (create a report, publish a publish, run a question)
- Sources — information the mannequin can learn (paperwork, rows, analytics)
- Prompts — reusable activity templates
Purchasers join over stdio for native servers, or SSE / streamable HTTP for distant ones. The distinction from a plain REST API is discovery: the mannequin introspects what the server affords and figures out the way to use it. You do not write per-task integration glue. Claude Code, Cursor, and Claude itself all communicate MCP natively, so one server makes your system operable by your total AI toolchain.
Hold each definitions loaded. The attention-grabbing half is how they compose.
Why whole-stack vibe coding fails
The frontend and backend fail in a different way, and vibe coding’s suggestions loop solely catches one sort of failure.
Frontend defects are seen and low cost. A misaligned card, a damaged button, a janky transition — you see it immediately, you screenshot it, you immediate the repair. The analysis technique of vibe coding (“does it look and behave proper?”) is a genuinely legitimate take a look at for presentation code.
Backend defects are invisible and costly. An auth verify lacking from one endpoint. A race situation in a stability replace. Row-level safety that was by no means enabled. These produce no visible sign, move the happy-path demo, and detonate weeks later with actual customers and actual information. “Does it appear to work?” is just not a legitimate take a look at for correctness, and correctness is the backend’s total job.
Each viral vibe coding horror story — leaked keys, deleted manufacturing information, wide-open Supabase tables — traces again to making use of a visual-feedback methodology to an invisible-failure area.
So do not. Cut up the stack alongside the failure-mode boundary.
Case examine 1: rebuilding a music platform in a couple of months
Our oldest product is a platform for musicians — 1000’s of artists, years of accrued legacy code from our earlier period as a 20-developer firm. This 12 months we rebuilt it finish to finish. It is reside at music.loop.followers.
The rebuild had three phases:
Part 1: API-first backend. Our dev staff (three engineers, led by our head of engineering) extracted the core area — artists, releases, fan relationships, rewards — behind clear, documented, open APIs. That is typical, cautious engineering. People wrote it, people reviewed it, people personal it. It is also the place our devs use AI closely, however in supervised mode: Cursor and Claude Code as pair programmers, each line learn earlier than merge. AI-assisted is just not the identical as AI-trusted.
Part 2: a light-weight MCP server over the APIs. As soon as the APIs existed, wrapping them in an MCP server was nearly trivial — a skinny layer mapping API operations to MCP instruments and browse endpoints to sources. We had Lovable generate the primary model of those wrappers. “Gentle” is deliberate: no enterprise logic lives within the MCP layer. It is a translation floor. All invariants keep within the backend the place they’re enforced as soon as.
Part 3: vibe coded frontend. Your complete user-facing software was inbuilt Lovable towards these APIs. Design iterations that used to take a dash took a day. When one thing regarded unsuitable, we screenshotted it and prompted the repair.
Complete elapsed time: a couple of months. The previous-generation estimate for a similar rebuild with the 20-person staff was roughly a 12 months — and the trustworthy anticipated worth was “by no means completed.”
Case examine 2: shopper product on blockchain rails
Our tourism product — a coalition loyalty move connecting ~20 companion venues throughout a regional vacation spot in Australia — pushes the sample more durable, as a result of the backend contains infrastructure that completely can’t be vibed: a customized backend and our personal blockchain pockets layer.
The pockets tech comes from our earlier lives. We constructed an NFT launchpad with self-developed embedded wallets, and earlier than that an AMM DEX on Terra that crossed $150M TVL (after which rode the Terra collapse to zero — a separate article’s price of classes). The purpose is: we all know precisely what it prices to get pockets infrastructure proper, which is why no AI writes it unsupervised.
And but the consumer-facing product — the move members carry, the rewards flows, the referral mechanics — is a completely vibe coded Lovable frontend sitting on these rails. Members get a quick, polished app. Beneath is infrastructure with years of hard-won engineering in it. Neither layer may do the opposite’s job.
Case examine 3: the build-vs-buy inversion
Here is a second-order impact no one warned us about: when frontend building value drops ~10x, your SaaS subscriptions cease making sense.
Most B2B SaaS you pay for is a CRUD frontend over commodity infrastructure, priced per seat. We have now changed, with vibe coded inner instruments:
- Brevo → our personal mail server and e mail platform
- Drip → our personal LinkedIn outreach software
- Our weblog/search engine optimisation stack → a customized CMS and publishing system
Every took days, not weeks. Every does precisely what we’d like and nothing else. Every killed a recurring invoice. The basic build-vs-buy argument — “your engineers’ time is price greater than the subscription” — assumed frontend building was costly. That assumption is gone, and I do not assume most SaaS pricing has observed but.
The MCP layer: the place it compounds
Individually, the inner instruments are a pleasant value story. Linked, they’re an working mannequin.
We had Lovable generate mild MCP servers over all of our inner APIs — weblog and search engine optimisation tooling, outreach, product information. Each one among our AI brokers can now function the corporate’s programs straight:
- An agent drafts, optimizes, and publishes a weblog publish by the CMS MCP server
- An agent manages outreach sequences by the outreach server
- An agent pulls reside product information to reply operational questions
No dashboards. No clicking by admin panels. No brittle per-integration glue code — the mannequin discovers the instruments and makes use of them.
The compounding loop seems to be like this:
- Vibe coding makes constructing a software low cost
- An open API makes the software programmable
- A lightweight MCP server makes the software AI-operable
- AI-operable instruments make the subsequent software cheaper to construct and run
Every product we ship makes the entire system extra automatable. That flywheel, greater than any particular person rebuild, is why 3 builders now outrun the outdated 20.
Failure modes we truly hit
So this does not learn like an advert, the issues that went unsuitable:
Context collapse on lengthy periods. Previous a sure dialog size, the AI begins contradicting its personal earlier selections — re-breaking fastened bugs, drifting from the design system. Repair: deal with periods as disposable. When one spirals, begin contemporary and re-state the spec. Immediate historical past is just not an asset; the deployed code is.
Silent schema mutations. Early on, we let the frontend software contact the database, and it “helpfully” altered tables to make a UI characteristic simpler. That was the day the rule grew to become absolute: the vibe layer consumes APIs, it by no means owns information. Supabase row-level safety plus API-only entry enforced it structurally.
Regression whack-a-mole. AI edits can repair the factor you requested about whereas quietly breaking an adjoining circulation, and visible evaluate misses it. We added light-weight end-to-end smoke assessments on vital paths (signup, redemption, sharing). Boring, efficient, non-negotiable.
The 80% lure. The primary 80% of a frontend arrives at miraculous pace; the final 20% — edge circumstances, empty states, offline dealing with — arrives at roughly regular pace. Funds accordingly, and do not demo the 80% as if it is the product.
The place a human nonetheless opinions every part: something touching auth, cash, or private information — even within the frontend. A senior dev audits these paths earlier than they ship. At some point of evaluate is the most affordable incident insurance coverage that exists.
The playbook
If you wish to run this sample:
- Draw the failure-mode boundary by your stack. Seen-failure code (UI, flows, presentation) goes to the vibe layer. Invisible-failure code (information, auth, cash, something with invariants) goes to engineers — AI-assisted in Cursor/Claude Code, human-reviewed, human-owned.
- Make the backend API-first, even when you’re the one shopper. Clear APIs are what make the frontend swappable and low cost. They’re additionally the prerequisite for step 4.
- Vibe code towards the APIs, by no means towards the database. Implement it structurally (RLS, scoped keys), not by coverage.
- Wrap your APIs in mild MCP servers. Skinny translation layers, zero enterprise logic. This can be a day of labor per service and it makes every part you have constructed operable by each MCP-speaking AI software you will ever undertake.
- Add smoke assessments on cash paths and auth paths. The vibe layer strikes too quick for handbook regression checking.
- Re-run build-vs-buy in your SaaS stack. Begin with the instruments which might be simply CRUD over your individual information.
What this implies for engineering groups
We reduce 20 builders to three out of necessity — the burn was going to kill the corporate — and found the smaller staff was quicker. I do not assume that generalizes to “fireplace your engineers.” I believe it generalizes to this: the scarce ability has moved up a degree. Writing code is plentiful now. Designing programs, drawing the boundary accurately, defining APIs, and understanding what beauty like — that is what our three engineers do all day, and so they’re extra leveraged than any staff I’ve ever run.
Vibe coding by itself builds demos. Vibe coding plus a hand-built system of report, open APIs, and an MCP layer builds corporations. The vibes are half the job — however solely half, and understanding which half is the entire recreation.