The Pirate Launch Kit
for Your Next Agentic SaaS
Next time you sit down to build, skip the 80% scaffolding every AI app needs anyway – and put your weekend into the part only you can build.
The 80% That Eats Your Weekend
You sit down on Friday evening with a full weekend ahead and a real product idea you've been turning over for weeks. By Saturday afternoon you check the clock. I should be done by now.
You're not. You're three hours into wiring Stripe webhooks. You haven't built the thing you came to build.
This is the weekend tax every vibe coder pays, and almost nobody talks about how big it is. It's not a bad day. It's not bad luck. It's 80% of the time, every time, on every new project. Auth flows. Database schemas. Payment webhooks. Deploy configs. Env-var roulette. The kind of work that has nothing to do with whether your idea is any good – but blocks every minute of the work that actually does.
I felt it cleanest at the Vibe Hackathon Cologne in March. Christina and I were building Bridesmaid – an AI wedding planner. Christina was working on what only she could shape: the agent's character, the specific tools we wanted to give brides, the visual language of the imagery the agent would generate for them. The product. The thing that makes Bridesmaid Bridesmaid.
I spent roughly 80% of the weekend on the scaffolding. Not because I wanted to. Because nobody has done it for us yet.
And then I noticed something worse. Even the part of the scaffolding that did feel like product work – wiring the agentic experience itself, getting streaming and tool use right – was going to be the same agentic core in every app I'd build after this one. The agent layer is the new CRUD. Once you see that, it stops being product work and becomes scaffolding too.
If the Pirate Launch Kit had existed, the hackathon would have gone where it should have gone: to the things only Bridesmaid needs.
I've watched the same dynamic play out in every Pirate Forge workshop I've run for the last nine months. While I'm there to teach people to vibe-code their app, most of the time goes to teaching them the scaffolding around it. Frustrating for me. More frustrating for them. They came to build their idea. They left with a slightly better understanding of webhook signatures.
Here's the part I want to say out loud:
Non-technical people have to pick up – from the past 50 years of the developer community – a habit of making things, in many cases, ridiculously overcomplicated. Why? Probably because they can. Somebody who is not coming from a place of terminal nostalgia and handcrafted lines of code looks at most of this stuff and is right to ask, what the fu…?
That's not a swipe at developers. It's pattern recognition. A lot of accidental complexity got baked in over fifty years because the people building the tools were the same people who could (and frankly enjoyed) jumping through every hoop. A non-technical builder showing up in 2026 inherits all of it whether they asked to or not.
And the tax doesn't end when the weekend ends.
A few weeks after the hackathon, Bridesmaid is alive. The shell is real – there's a domain, a sign-in, a payment flow. The functionality is still too thin to confidently put a marketing campaign behind. We could have used the hackathon time to build that depth – if the shell hadn't taken the whole weekend. Now we're stealing an hour here and an evening there from real life to push it forward, and honestly, it is tough.
That's the cost the kit is really attacking. Not the 48 hours. The weeks of we'll get there that follow them – the weeks that almost never quite arrive.
The New SaaS Shape
If the tax is the symptom, the shape is the disease. The shape we inherited.
For most of SaaS history, “the software” was a workspace. You opened it, and you did the work in it. PowerPoint on your desktop. Google Slides in the browser. The leap from PowerPoint to Google Slides was real – collaboration, no-install access, version control without filenames – but the philosophy didn't move. The tool gave you the surface. You did the work.
Then Gamma showed up. You describe what you want. You drop in some text context. Gamma builds the deck. You don't open a slide and click “insert text box.” You look at what Gamma made, you say I like this, change that, the title is wrong, and Gamma changes it. Roughly 90% of the work has shifted off you. The 10% that stayed is the part that's actually you: taste, direction, approval, the call to redo this slide because it's not landing right.
Watch what happened to the UI. It stopped being a workspace. It became a supervision surface – the place where you brief the agent, review the output, and steer the next step. Same screen, completely different job.
Claude Design made the pattern even clearer to me: one central agent, working with full context from an associated codebase, doing things like producing a presentation while you look on and direct. The agent is the engine. The UI is the bridge.
I believe this is the future shape of SaaS, and the kit is built for it:
The agent does the work. The SaaS coordinates the business. The UI is where the human supervises.
When the agent is the thing doing the work, the pricing model shifts naturally. Outcome-based and usage-based billing stop feeling exotic – they're the obvious match for a business where the value is the work delivered, not the seat with access to a tool. We're already seeing this with Gamma and Claude Design. It's likely where most products worth building are headed.
The objection that always comes up:
Isn't this just a slop factory? Won't agents producing 90% of the work flood the world with mediocrity?
Real concern. The answer isn't to hand it all to the agent and walk away. The answer is human-in-the-loop at the key moments – feedback gates, approval steps, the places where taste actually matters. The aim of a well-built agentic SaaS is a process that improves with every cycle: less oversight required, more throughput, stronger consistency, while preserving the human taste with which it was instructed in the first place. The human's taste is the seed. The human's approval is the gate. The agent's job is everything in between – and to get incrementally better at needing the human less for the routine work, more for the decisions that matter.
A starter kit built for this shape is not a starter kit built for old SaaS with an AI bolt-on. The pieces are different. The defaults are different. The runway is different.
Here's what's in it.
The Pirate Launch Kit
Four layers. One opinionated stack.
The Frontend Layer
The human supervision surface – where users brief the agent, inspect progress, approve outputs, correct mistakes, and steer the next step. Not “just the UI.” The command bridge.
Next.js 16
What it is: The frame your app lives in. Next.js handles routing, server-rendered pages, API endpoints, and the structure that lets one project be both your landing page and your logged-in app. It runs natively on Vercel, which is where this kit ships. Docs →
Why this choice: AI tools know Next.js inside out, the App Router is the current default, and it's the path of least resistance from empty folder to live URL.
AI assist: Official Next.js MCP for live, version-matched context. The nextjs skill is bundled in the Vercel plugin. Next.js MCP →
Price: Free and open source — use it forever, no account needed.
Tailwind CSS + shadcn/ui
What it is: Tailwind is how you style things — write classes directly in your markup instead of juggling stylesheets. shadcn/ui gives you copy-pasteable components (buttons, dialogs, forms, dropdowns) that live inside your repo, so you can edit them like normal code instead of fighting a black-box library. Docs →
Why this choice: Tailwind is the styling syntax AI agents are most reliable with, and shadcn/ui means your design system is ordinary code in your project, not a sealed dependency.
AI assist: Official shadcn MCP for browsing, searching, and installing components live. The shadcn skill is bundled in the Vercel plugin. shadcn MCP →
Price: Both free and open source — Tailwind costs nothing, and shadcn/ui components live in your repo as ordinary code.
AI Elements
What it is: A component library on top of shadcn/ui for the agent-facing parts of your app — chat surfaces, prompt fields, response views, reasoning panels, source citations, tool-call displays. The pre-built pieces of the supervision UI. Docs →
Why this choice: We shouldn't rebuild the same agent conversation, reasoning, source, and response components in every product.
AI assist: Bundled in the Vercel plugin via /vercel-plugin:ai-elements.
Price: Free and open source — same copy-into-your-project model as shadcn/ui.
Playwright
What it is: A real browser your agent (or a test) can drive. Two jobs: end-to-end tests for the flows that matter (signup, login, payment, the core journey), and live browser automation — your agent opens your app, takes screenshots, clicks around, and reports back on what looks broken or off. Docs →
Why this choice: For vibe coders, the most useful test isn't “ does every internal function pass?” but “can a real user complete the core journey?” And just as valuable: handing your agent eyes on the screen so it can catch obvious mistakes before you do.
AI assist: Coding-agent CLI and installable skills mean your agent can open the browser, navigate your app, take screenshots, and suggest fixes — not just write tests. Coding agents CLI → · Playwright Skills →
Price: Free and open source — Microsoft maintains it, no fees ever.
The Backend Layer
The SaaS operating system – identity, data, billing, email, analytics, files, deployment. The boring infrastructure that must work before the magic interaction matters. The goal isn't maximum flexibility – it's fewer decisions, fewer accounts, fewer ways to break production.
Vercel
What it is: Where your app lives once it's online. Vercel handles deploys, preview URLs for every change, custom domains, scheduled jobs, and file storage. Calls itself an “AI Cloud” because it's tuned for Next.js, AI workloads, and agent backends out of the box. Docs →
Why this choice: It lets you deploy fast without learning infrastructure, and keeps your app, storage, workflows, and AI models in one platform.
AI assist: Official Vercel MCP for live access to docs, projects, deployments, and logs. Deeper guidance comes through the Vercel plugin. Agent resources →
Price: Free Hobby tier covers 1M requests + 100GB bandwidth/month — roughly a small side-project getting a few thousand visitors a day before you'd hit any limit. $20/user/mo Pro when you outgrow it.
Clerk
What it is: Authentication done for you — sign-up, sign-in, sessions, password resets, social login, organizations, team invites, role-based access. Polished UI components included. Docs →
Why this choice: Auth is too risky to vibe-code from scratch. Clerk gives non-technical builders safe defaults and good-looking components without learning the entire auth domain first.
AI assist: Official remote MCP server with current SDK snippets and implementation patterns, plus official Clerk Skills your agent can install and follow. Clerk MCP → · Clerk Skills →
Price: Free up to 50,000 monthly active users (basically your first 50k unique sign-ins per month — plenty for most launches). $25/mo Pro after that.
Neon + Drizzle
What it is: Neon is your hosted Postgres database — fast, serverless, no Docker, no local setup. Drizzle is the TypeScript layer on top: define your tables in code, query them with autocomplete, run migrations safely. Docs →
Why this choice: Vibe coders shouldn't be running databases on their laptop. Neon is remote-first by design, and Drizzle keeps the schema and queries readable both for you and for AI agents.
AI assist: Official MCP server for managing projects, branches, databases, queries, and schemas. Neon also publishes AI rules with specific Neon + Drizzle patterns for tools like Cursor. Neon AI rules →
Price: Neon free tier gives you 0.5GB storage + 100 compute-hours/month — enough for a small app database with a handful of daily users. Pay-as-you-go after that. Drizzle is free and open source.
Stripe
What it is: Subscriptions, checkout, billing portals, invoices, one-off payments, and later usage-based billing. The default way SaaS gets paid. Docs →
Why this choice: Billing isn't where we want to improvise. Stripe is the most battle-tested SaaS payment path with the strongest ecosystem.
AI assist: Official MCP server for the Stripe API and docs, plus a Stripe Agent Toolkit for Vercel AI SDK and other agent frameworks. No first-party Stripe skill yet. Stripe MCP → · Building with AI →
Price: No monthly fee. Stripe's headline rate is 2.9% + $0.30 per successful card charge — so on a $100 sale you keep about $96.80. Lower in some regions, higher for international cards.
Resend
What it is: Email infrastructure for your app. Start with transactional emails — sign-up confirmations, receipts, password resets, system notifications, and anything your agent needs to send. Resend also has a separate newsletter product (Resend Broadcasts) you can add later when you actually want to talk to your whole user base. Templates are React components instead of fragile HTML. Docs →
Why this choice: Email should be boring, modern, and easy to wire into a Next.js app. Resend has the cleanest founder-friendly DX right now.
AI assist: Official MCP server, plus official skills for sending emails, building React Email templates, using the Resend CLI, and agent inbox workflows. Resend MCP → · Resend Skill → · CLI for agents →
Price: Free up to 3,000 emails/month — enough for the transactional emails of a starting product. $20/mo Pro for 50,000 emails after that.
PostHog
What it is: What your users actually do, in one place — web and product analytics, funnels, session replays, feature flags, A/B experiments, surveys, and lightweight error tracking. Docs →
Why this choice: Lets a founder understand what users do, where they get stuck, and which experiments work without starting with five separate tools.
AI assist: Official MCP server for analytics queries, feature flag management, and error investigation. PostHog also has a skills store for team-specific analytics skills, plus PostHog Wizard — an agentic CLI that installs and configures PostHog in your codebase. PostHog MCP → · Skills store →
Price: Free up to 1 million events/month (an event = any tracked click, pageview, or action — 1M is roughly a few months of solid analytics for a small launch). Pennies per event after that.
The Agent Layer
Where the product stops being a dashboard and starts doing work: model calls, tool calls, streaming, structured outputs, durable execution, external agent channels. The part that makes the SaaS agentic-native instead of old-school CRUD with an AI wrapper.
Vercel AI SDK
What it is: The TypeScript framework for talking to AI models from your app. One consistent API for streaming responses, tool calls, structured outputs, and switching providers — built natively into the React + Next.js setup this kit uses. Docs →
Why this choice: Keeps AI implementation inside the TypeScript world instead of scattering raw provider SDK calls across your codebase.
AI assist: Supports MCP tools directly. Bundled in the Vercel plugin via /vercel-plugin:ai-sdk.
Price: Free and open source — you only pay for the AI model calls themselves.
Vercel AI Gateway
What it is: One endpoint, hundreds of models. Set budgets, monitor usage, balance load between providers, and fall back automatically when one is down. Docs →
Why this choice: Prevents model-provider chaos by giving the app one place to manage access, costs, observability, and fallback behavior.
AI assist: Covered by the Vercel plugin's /vercel-plugin:ai-gateway skill — no separate setup needed.
Price: Free $5/month credit included. After that, pay-as-you-go at the model's exact list price — Vercel adds zero markup.
Workflow SDK
What it is: Makes long-running async functions in your app durable. Multi-step jobs that need to retry, sleep, hold state, and survive failures — research flows, document generation, onboarding sequences, agentic back-office work. Docs →
Why this choice: Real agent work isn't a single API call. It needs steps, retries, sleeps, and recovery without you having to run queues and workers.
AI assist: Bundled in the Vercel plugin via /vercel-plugin:workflow — covers build, init, and migration.
Price: Free and open source. If you run it on Vercel, you just pay normal Vercel compute costs — no separate workflow fee.
Vercel Chat SDK
What it is: Build agents that show up in Slack, WhatsApp, Discord, Teams, Linear, GitHub, and other places your users already work — not only inside your web app. Docs →
Why this choice: Lets the same agentic product meet users where they already are instead of forcing every interaction into the web UI.
AI assist: Bundled in the Vercel plugin via /vercel-plugin:chat-sdk.
Price: Free and open source — the SDK itself costs nothing. You'll pay whatever each channel (Slack, WhatsApp, etc.) normally charges for its own API.
The Builder Layer
The vibe-coding development experience – the tools, skills, MCP servers, CLIs, and companion guidance that help a non-technical builder wire the stack, set env vars, deploy, and iterate fast. Not the product architecture. The launch mechanic.
Pick one AI coding assistant first. My primary recommendation is Claude Code + VS Code (below). Cursor, OpenAI Codex, and Google's Antigravity are equally fine alternatives – pick what you're comfortable with. Whichever you choose, the rest of this layer works the same.
Claude Code
What it is: Anthropic's AI coding assistant. Reads your whole codebase, edits multiple files, runs commands, and automates parts of your build flow. Available as a CLI you can run from any terminal or as an extension inside VS Code. Docs →
Why this choice: The recommended default builder workflow — gives non-technical builders a strong agentic coding loop without making the product itself dependent on it.
Price: Included in the Claude Pro plan ($20/mo). Heavier users go to Max starting at $100/mo for 5× the usage.
VS Code
What it is: Microsoft's free code editor and the recommended host for Claude Code. The official Claude Code extension adds inline diffs, file mentions, plan review, and keyboard shortcuts directly inside the editor. Docs →
Why this choice: Free, familiar, and the editor most coding agents and tutorials assume. Pairs cleanly with Claude Code without locking you into an AI-specific IDE.
Price: Free and open source — Microsoft maintains it.
Vercel Plugin
What it is: One install (npx plugins add vercel/vercel-plugin) drops 25 curated skills, 3 specialist agents, and 5 slash commands into your coding assistant – covering Next.js, shadcn, AI SDK, AI Gateway, Workflow SDK, Chat SDK, env vars, deploys, and more. Works in Claude Code and Cursor today; requires Bun (not just Node). For deeper guidance, invoke a specific slash command like /vercel-plugin:ai-sdk or /vercel-plugin:workflow. Docs →
Why this choice: Most of the Vercel-family AI assistance in this kit flows through this single install. Replaces a half-dozen separate skill installs with one foundational setup. Auto-updates as Vercel ships new skills.
Price: Free. Telemetry on by default — set VERCEL_PLUGIN_TELEMETRY=off to disable.
MCP servers — an open standard that lets your AI assistant talk directly to outside tools and data, so it asks the live system for current state instead of guessing from training data. Each tool above publishes its own.
Skills — packaged instructions, scripts, and workflows that teach an assistant how to do a specific task reliably. Anthropic, OpenAI, Cursor, and Vercel all ship them in compatible formats.
CLIs and registries — the plain commands the assistant runs for you to initialize, install, scaffold, verify, and deploy.
Scaffold Your Next One in Minutes
You've read all this. What does it actually look like?
You make a new folder for your project, open your AI coding assistant inside it, and paste in the install prompt. That's the entire setup. From here the skill drives.
First it calibrates. How deep is your technical knowledge on a scale of 0 to 10? That answer changes how it talks to you for the rest of the session – a 2 gets every term explained inline, a 9 gets terse commands. Then it recommends auto-approve mode so it can move fast and only stop you for things that cost money or touch your accounts. You pick.
Then it builds the foundation. Checks your folder is empty and named in a way npm will accept. Installs whatever runtimes you're missing (Node, Bun, git, the GitHub / Vercel / Stripe CLIs) itself, never telling you to “go download” anything. Initializes git with a pre-push hook that blocks pushes containing secret-shaped strings. Loads the Vercel plugin so it has up-to-date patterns for Next.js, shadcn, AI Elements, AI SDK, AI Gateway, Workflow SDK, and Chat SDK. Writes a live CHECKLIST.md it ticks off in real time so you always know where you are.
Then it ships you something live, fast. About 30 minutes in, you have a deployed *.vercel.app URL with the agent shell running on it – light/dark theming wired, Geist fonts, polished chat UI from AI Elements, streaming responses from Claude. There's a one-time password gate in front so you can share the URL with friends before Clerk signup is even wired. The magic moment – you typing into a chat on your phone and seeing it stream back – lands at the 30-minute mark, not the 90.
Then the backend, one service at a time. Clerk swaps the password gate for real signup. Neon + Drizzle for the database with a lazy-sync helper that creates the user row on first request. Resend for the welcome email. Stripe for test-mode checkout with a webhook handler that verifies signatures. PostHog for analytics. After every service the skill redeploys and smokes the live URL with you – you watch the same URL get more capable each time. No copy-pasted API keys: keys flow through Vercel Marketplace integrations or get pushed via vercel env add over stdin.
Then production cutover. Custom domain wired. Clerk on a production instance. Stripe in live mode with a restricted key (not the root secret) and one real €0.50 charge that the skill refunds for you. Resend on a verified sending domain so the welcome email lands in inbox, not spam. PostHog on a production project. The skill pauses and confirms before each irreversible step – this is the only phase where it slows down.
The whole time, it asks you only for what only you can do. Account creation. Domain choice. The moments where a real human signs into a real third-party dashboard. Installs, env vars, deploys, smoke tests, secret hygiene – the skill handles all of it, asking permission before anything that touches money or your accounts.
At the end you have a live custom domain with landing, signup, a working chat that streams Claude, a real database under it, real welcome emails going out, real payments working, and real analytics flowing. The runway is built. The clock you've been running against hasn't even started yet.
From here, everything that's actually you starts. The agent's character. The specific tools you want it to have. The visual language. The magic interaction only your product needs. None of that is the kit's job – the kit's job is to hand you the runway and step out of the way.
The honest version of our hackathon weekend with Bridesmaid is sitting at the top of this article. The shell took the whole weekend. The kit would have given the weekend back. What we'd have done with that weekend – the depth, the agent character, the imagery for brides, the actual product – is the part we're still squeezing in from real life, week by week. That's the cost the kit is really attacking.
Install the Pirate Launch Kit skill
First, create a new empty folder for your project and cd into it – the kit installs into the current folder so it only loads when you're working on this project. Then pick one of the two install paths below.
Option 1 — Run it in your terminal
From inside your new empty project folder:
$ curl -fsSL pirateskills.com/launch-kit/install | shOption 2 — Hand it to your AI coding assistant
Step 1. Open your AI coding assistant (Claude Code, Cursor, Codex, Antigravity) inside a new empty folder for your project, then paste this:
Install the Pirate Launch Kit skill into this project by running:
curl -fsSL pirateskills.com/launch-kit/install | sh
The install drops the skill into this project's .claude/skills/ folder. When it's done, tell me to start a fresh conversation in the same folder so the new skill becomes available, and remind me to paste the next prompt I have.Step 2. Start a new conversation in the same folder, then paste this:
Use the pirate-launch-kit skill to scaffold a new agentic SaaS project from scratch in this empty folder. Walk me through it step by step. Ask me only for the things you genuinely need from me — account creation, API keys, copy decisions — and handle everything else (installs, env vars, deploys, smoke tests) yourself. Confirm with me before any major action.By the end, you have a live URL with a landing page, login, and streaming chat shell — ready to put your actual product on top of.
Build the runway in minutes.
Spend the weekend on the part only you can build.
Build. Grow. Repeat.
Cheers,
Ben
Ready to Go Deeper?
Builder Lab
Free weekly online session where we build real projects together. Bring your idea, install the kit, and ship the shell live.
Builder Forge
·In 58 days · Wed, Jul 112-week cohort program to build, ship, and grow your agentic SaaS. The Pirate Launch Kit is your hour-zero starting point.
Growth Lab
Free weekly online session where we work on funnels and positioning together. Bring your shipped shell and grow it.
Growth Forge
·In 9 days · Wed, May 136-week cohort program to build predictable, scalable revenue around the product the kit gave you the runway to build.
Questions & Answers
Founder from Cologne with 15 years of startup experience across 9 ventures. After helping thousands master growth marketing, Ben learned vibe coding from scratch and launched CaptAIn within three months. He leads the Vibe Coding Cologne community, blending real founder experience with teaching clarity.