Fourteen tools, arranged so that an AI can do most of the typing. Here is what each one is actually for, in plain language — and the two places where the word "free" will cost you money.
- The stack does six jobs: build the app, write the code, store the data, call the models, ship it, watch it. Fourteen tools, six jobs.
- Two AI coders, not one. The split is deliberate — one is set on the interface, the other on the logic.
- Vercel's free plan does not allow commercial use. Their own words: it "restricts users to non-commercial, personal use only". Ads, payments, affiliate links or client work all require Pro at $20/user/month. The most expensive misunderstanding in the stack.
- Railway’s "Free" plan gives $1 a month of credit — far less than it costs to keep a service running. Budget for $5/month from the start.
- You do not need all fourteen to start. Four of them get a working app in front of a person. The rest are for when something is already live.
- Nothing here is required. Every layer has alternatives, and the stack is opinionated in ways worth knowing before you commit.
This is a stack that circulates widely on social media as a list of tools for building with AI. Lists like it are useful as a starting map and unreliable as a source of fact — they are written to be saved, not checked.
So the tools are described here, and the claims attached to them are verified separately. Where a claim turned out to be wrong, this page says so and gives the checked version. Where a figure could not be confirmed, it is not repeated.
The six jobs, not fourteen tools
A list of fourteen names is hard to hold. The same list grouped by what each thing does is much easier, and it also shows you where the gaps and the overlaps are.
Next.js · Tailwind2
Claude Code · Codex2
Clerk · MongoDB · Pinecone3
OpenRouter1
GitHub · Vercel · Railway3
Sentry · Linear · Featurebase3
Layers 1 to 4 build the thing. Layer 5 puts it online. Layer 6 only matters once real people are using it — and adopting it early is the most common way to spend a week on tooling instead of on the product.
1 · The frame
Next.js — the shape of the app
A framework is a set of decisions already made for you: how pages are organised, how the address bar maps to what you see, how data reaches the screen. Without one, you make all of those choices yourself, badly, in the first week.
Why this one: Next.js is built and maintained by Vercel, the hosting company in layer 5, so deploying there is close to frictionless. That convenience is also the tie-in — two of the fourteen layers are the same vendor, and the stack is shaped around it. Next.js runs elsewhere, but you give up some of the integration you chose it for.Next.js is maintained by Vercel; see vercel.com/docs. Checked 26 Aug 2026
Skip it if: you are building something with no pages — a script, a bot, a data job. A framework is overhead you will not use.
Tailwind — how it looks
Instead of writing style rules in a separate file, you write short class names directly on the thing you are styling. text-lg font-bold rather than inventing a name, opening another file, and defining it there.
The real reason it is in an AI stack: the styling lives next to the markup, so a model editing one line can see and change the appearance in the same edit. Nothing has to be kept in sync across two files.
The honest cost: your markup gets long and noisy. People either stop noticing within a week or never make peace with it.
2 · The builders — and why two
The stack uses two AI coding tools rather than one, which looks like duplication until you see the split.
Two models, two jobs
One on the interface — layout, spacing, how a screen feels, making something look finished rather than like a demo.
One on the logic — the parts with no visual output: data handling, edge cases, the bug that only appears on the third attempt.
Is the split real? Partly. Models genuinely differ in temperament, and anyone using several daily forms preferences. But "best at design" and "best at logic" are impressions from use, not measurements, and the ranking changes with almost every release. Treat it as a reason to try two, not as a fixed fact about either.
The transferable idea outlives whichever models are current: run more than one, and give each the work it is better at. The cost of a second subscription is small against the cost of being stuck.
3 · The plumbing
Clerk — logins, without building logins
Sign-up, sign-in, password resets, "continue with Google", sessions, and the security around all of it. This is weeks of work with real consequences if you get it wrong, and it is completely undifferentiated — nobody chooses your product because of its login screen.
The free tier is genuinely large: 50,000 monthly retained users per app, on Clerk’s own pricing page.
Retained is not active, and the difference is in your favour. Clerk’s definition: a user only counts "if they return to your app at least 24 hours after signing up, so someone who signs up and never comes back does not count." Most competitors bill on monthly active users, which counts that person. Paid plans start at $25 a month billed monthly, $20 billed annually — and carry the same 50,000 allowance, so you upgrade for features, not headroom.Clerk pricing page, clerk.com/pricing, checked 26 Aug 2026. Reports elsewhere date the increase from a previous 10,000-user tier to early February 2026; that date is not stated on Clerk’s own page and is not asserted here.
MongoDB — where the data sits
A database that does not make you define the shape of your data up front. You can save a record with five fields today and seven tomorrow without a migration.
Why that suits AI-assisted building: early on the shape changes constantly. A database that objects to every change slows the loop you are relying on.
The counter-argument, which is real: the thing that protects you from mess later is exactly the strictness this avoids. Many teams start flexible and end up enforcing structure in their own code — doing the database's job manually. Flexible is the right default for finding out what you are building; it is not obviously right once you know.
Pinecone — memory for the AI part
The only layer that exists purely because there is AI in the product. A language model knows nothing about your documents. A vector database stores your material in a form that can be searched by meaning rather than by keyword, so the right three paragraphs can be found and handed to the model before it answers.
This is the machinery behind almost every "chat with your documents" feature. Without it, the model either does not know your content or you paste all of it into every request, which is slow and expensive.
Free tier, from Pinecone’s own limits page: 2 GB of index storage per organisation, 5 indexes per project, 100 namespaces per index, 1 project and 2 users per organisation, with 2 million write units and 1 million read units a month.Pinecone database limits reference, docs.pinecone.io, checked 26 Aug 2026. How many records 2 GB holds depends entirely on your vector dimensions, so no record count is given here — an earlier version of this page said "roughly 300,000 records" and "community support only", neither of which appears in Pinecone’s documentation.
Skip it entirely if your product does not answer questions about your own material. Most apps do not need one.
4 · OpenRouter — one key for every model
Why a middleman
Every AI provider has its own account, its own key, its own bill, its own way of being called. OpenRouter sits in front of them: one key, one bill, and you change which model you are using by changing a string.
What that actually buys you is not convenience. It is the ability to change your mind. Model rankings move constantly — a provider that was the obvious choice in spring is mid-table by autumn. Wiring your app directly to one vendor makes switching a rewrite. Going through a router makes it a config change.
The trade: another company between you and the model, taking a margin and able to have an outage of its own.
5 · Shipping — and the expensive misunderstanding
GitHub — the undo button
Version control keeps every state your code has ever been in, so any change can be reversed. This matters more with AI in the loop, not less: when a model edits twelve files and something breaks, the question is not "what did it change" but "how do I get back". Without version history, you cannot.
It also triggers deployment — push, and the site updates itself.
The claim attached to Vercel in stack lists is usually "free until you scale". That is the wrong axis. The Hobby plan is restricted to personal, non-commercial use — the limit is not how big you get, it is whether you make money.
Ads, payments, affiliate links, or work delivered to a client all fall outside it. The moment your side project earns anything, you need the Pro plan at $20 per user per month.
Plenty of people learn this after building on Hobby for months. If there is any chance the thing will earn money, the correct number to plan around is $20 a month, not zero.
Vercel documentation, /docs/plans/hobby and /docs/pricing, checked 26 Aug 2026. Their exact wording: "the Hobby plan restricts users to non-commercial, personal use only." Hobby also carries usage limits — 1 million edge requests, 1 million function invocations, 4 CPU-hours, 200 projects and 100 deployments a day — but the commercial-use restriction is the one that catches people.Railway’s original free tier was withdrawn in 2023. A plan called Free still exists, which is where the confusion starts — but it carries $1 of monthly usage credit, and Railway prices RAM at $10/GB/month and CPU at $20/vCPU/month. A dollar does not run a service for a month.
New accounts get a one-time $5 trial grant. The realistic plan for anything that stays up is Hobby at $5 a month, which includes $5 of usage credit — so you pay the $5 whether you use it or not, and more if you exceed it. Treat Railway as a paid line item from the start.
Railway's job in the stack is real: Vercel is built for front ends, and some work does not fit that shape — background jobs, scheduled tasks, long-running processes, anything that needs to keep running when nobody is looking at the page. Railway runs those. But budget for it from day one rather than discovering it.
Railway documentation, docs.railway.com/reference/pricing/plans, checked 26 Aug 2026: Free $0/mo with $1 credit · Hobby $5/mo with $5 credit · Pro $20/mo with $20 credit · one-time $5 trial grant. An earlier version of this page said Railway "has no free tier" and gave a 30-day trial window with no card required — the first is too strong and the second is not stated in Railway’s own documentation. Both corrected.6 · After it is live
Three tools for three questions
What broke? An error tracker records crashes as they happen, with the exact line and the conditions. The alternative is finding out when a user bothers to tell you, and most never do — they leave.
What am I doing next? An issue tracker is one list of everything wrong or wanted, instead of notes across four places. The AI-specific angle: a coding agent can read and update the list, so the plan and the work stay attached.
What should I build? A feedback board lets users request things and vote, so you build against demand instead of guesswork.
All three are premature before you have users. An error tracker with no traffic reports nothing. A feedback board with no visitors is an empty room. This layer is the reward for shipping, not a prerequisite for it.
What you can skip, and in what order to add things
Fourteen tools is a finished setup, not a starting point. Adopting all of them before you have built anything is a well-worn way to spend a week configuring instead of making.
Each layer should be a response to a problem you actually have. If you cannot name the problem a tool solves for you this week, you do not need it this week.
What this stack is good for — and what it is not
Every tool here is a dependency. Fourteen tools is fourteen accounts, fourteen ways to be locked out, fourteen pricing pages that can change, and fourteen services that can have a bad day. The stack is genuinely fast to start with — that is not marketing — but the speed is borrowed against a maintenance surface that only grows.
The version of this advice worth keeping is not the list of names. Names expire; two of the fourteen here already carry claims that were true when written and are not now. What survives is the shape: know which job each layer does, add a layer only when you have the problem it solves, and check the pricing page yourself before you build on a number somebody put in a graphic.
Every figure on this page comes from the vendor’s own documentation, not from a summary of it: vercel.com/docs/plans/hobby and /docs/pricing · docs.railway.com/reference/pricing/plans · clerk.com/pricing · docs.pinecone.io database limits reference. All checked 26 August 2026.
Four figures in the first version of this page came from third-party summaries and did not survive checking against the vendors. They are corrected above and the wrong versions are named, because that is the same failure this page warns about.
Pricing and free tiers move faster than any article. Check the vendor's own page before you commit — that is the whole lesson of the two corrections above. Errors here are logged at corrections.
The stack is a reasonable set of defaults. The claims that travel with it are not, and the two that cost money were both about the word "free".