"Software for one" — apps built for an audience of exactly one household, not a market — is having its second moment on Hacker News. Alex Waxman's essay describing four personal apps he built with Claude Code hit 105 points and 92 comments, reviving a six-year-old idea from writer Robin Sloan at the exact moment AI coding agents made it practical at scale. If you're weighing whether to build your own instead of paying for (or tolerating) an app that almost fits, this post breaks down the actual stack, the real monthly cost, and the strongest pushback from the discussion.
TL;DR — the software-for-one thesis
| Question | Answer |
|---|---|
| What changed since 2020? | Robin Sloan's original BoopSnoop app took a week, "half of it lost to code-signing purgatory." AI coding agents cut that to a single evening for comparable apps in 2026. |
| What's the reported stack? | Next.js/Vercel, Tailwind + shadcn/ui, Better Auth, Postgres on Neon + Drizzle, Resend, GitHub Actions, Claude Code + Vercel AI SDK |
| What's the monthly cost? | ~$160/mo for a builder running several projects; realistically $20-30/mo for one or two apps if you already pay for an agentic coding plan |
| Is code quality a concern? | Less than for shipped products — a HN commenter (pornel) notes the bar for "good enough" drops sharply when there's exactly one user who can just ask for a fix |
| What's the security risk? | Real but manageable — fewer third-party dependencies than typical apps, but a personal device is still not a sandbox, per HN pushback |
| Managed hosting or a VPS? | Split opinion — managed services (Vercel/Neon) are faster to start; a $5-15/mo VPS (Hetzner, DigitalOcean) is cheaper over the app's full lifetime |
| Does this need a developer's stack? | Today, mostly yes. Sam Altman's one-paragraph ChatGPT prompt building a full trip-coordination site suggests that requirement won't hold for long |
The four apps, and what they actually do
Waxman's essay walks through a sleep-tracking app that turns a sleep consultant's PDF — full of "wake windows, nap caps, what to do when a nap fails" — into a live, shareable schedule for him, his wife, and their nanny. A fitness app sizes his morning smoothie's eight ingredients to that day's run distance, pulling data from Strava, Oura, Withings, and FatSecret. A marathon plan derives heart rate zones from his actual race history instead of an age-based formula. A jazz quiz app tests him on chord voicings from his piano lessons.
None of these apps are technically hard. What changed is that building all four, well enough to trust with his family's daily routine, took a week of evenings combined — not months.
Where the idea comes from
The intellectual lineage here predates AI coding entirely, and it's worth knowing if you're building in this space. Clay Shirky's 2004 essay "Situated Software" described exactly this pattern two decades early: small, social-context-specific software written for a known group of users, deliberately opting out of the generality that "real" software products need. Robin Sloan's 2020 BoopSnoop post picked up the same thread with a concrete example — four downloads, and he called it a resounding success, because scale was never the goal.
What's new in 2026 isn't the idea. It's that the cost to execute it collapsed. Thariq and Vercel's Lee Robinson both made this point on X around the same time Waxman's essay resurfaced: personal computing can now be genuinely personal, without the app needing "user profiles, badges, subscription tiers" or any feature that isn't specifically for the person using it.
What Hacker News actually built, beyond the essay
The comment thread is arguably more useful than the essay itself, because it's a second data point from a different builder. tptacek listed a set of desktop tools built "in the last month or so": a menubar Apple Music player running an embedded agent loop that keeps a non-repeating "woodworking-appropriate" playlist full; a calorie tracker that resolves one-line meal descriptions ("oyakadon 1/2c rice, 2 thighs...") to macros via GPT-5 calls into SQLite; a combined Apple TV/Roku menubar remote; a LaTeX-rendering SageMath calculator front-end for linear algebra; a "self-driving wiki" that digests PDFs; and a picture-frame layout tool that outputs cut lists and rabbet depths.
His calorie tracker drew the most scrutiny in the thread — commenters questioned whether an LLM can estimate macros accurately from a one-line description. tptacek's response is a useful data point for anyone building something similar: he validated the model's output against all three frontier models, manually researched his most common recipes, and reconciled months of estimates against actual scale-weight loss. The key simplifying assumption he flagged: this works because he cooks almost all his own meals from known ingredients, and would be far less reliable for restaurant or packaged food without a web-search tool bolted on.
The strongest pushback, and whether it holds
Three objections came up repeatedly, and they're each worth weighing before you start building.
"A personal device is not a sandbox"
Commenter _dark_matter_ raised the direct security concern: running a pile of AI-generated code with real data access, on a phone or laptop, is still a real attack surface — "it just takes one bad library." tptacek's counter was that his tools are simple SwiftUI apps with minimal dependencies, and that AI-generated code tends to pull in fewer third-party packages than human-written equivalents, since an agent will often rewrite a small utility rather than add a dependency for it. Both points are true and don't fully cancel out — fewer dependencies reduces supply-chain risk, but doesn't eliminate the need for basic hygiene (no exposed secrets, minimal permissions, dependency scanning) even on a single-user app.
Quality bar: feature, not bug
pornel's framing is the sharpest one in the thread: single-user software mitigates the actual problem with LLM-generated code, which is inconsistent quality. A professional app has to handle every user's edge cases in a way other developers can maintain. A personal app only has to work for one person, and can be thrown away and rebuilt when it gets messy — which Waxman does explicitly with his sleep app, retiring it after four months once his son started sleeping through the night. One counter-argument (eviks) is fair too: discarding and rebuilding isn't free, it's just cheap enough now not to sting.
Managed hosting vs. a VPS
rrvsh argued for deploying to a cheap VPS (Hetzner, DigitalOcean) instead of stacking Vercel + Neon + auth-provider subscriptions, on the logic that $5-15/month for the app's full lifetime beats managed pricing that can change under you. Another commenter confirmed running three Hetzner VPS instances for about $14/month combined. This is a real trade-off, not a clear win either way — a VPS means you own patching, backups, and uptime yourself, while managed services trade a higher recurring cost for near-zero operational overhead. If you're already comfortable with a full AI coding agent stack, the VPS route isn't much harder to set up.
Distribution friction: the App Store problem
The thread also surfaced a real limitation Waxman's essay doesn't dwell on: getting a personal app onto an iPhone is still friction-heavy. Apple's $99/year developer fee, weekly TestFlight expirations on free accounts, and Siri's closed extension APIs were all cited as reasons iOS is the worst platform for this pattern. Alternatives discussed: PWAs (skip the App Store, lose some native APIs), WebXDC apps distributed peer-to-peer through Delta Chat's encrypted relay network, or just accepting the $99/year and a small cron job to handle re-signing. None of this applies if your personal app is web-based, which is why most of the stack examples in this piece — including Waxman's — default to a Next.js web app rather than a native one.
What this means if you're learning to build with AI agents
This pattern is a genuinely good on-ramp if you're newer to building with Claude Code — the stakes are low (you're the only user), the feedback loop is immediate (you use it the same day), and the debugging loop tptacek and Waxman both describe — "a feedback screenshot shared with Claude gets the job done 9/10 times" — is exactly the skill that transfers to building anything larger. It's also directly relevant to the broader argument, made elsewhere on Hacker News, about whether programming itself is becoming lower-skill work as agents absorb more of the mechanical part of building — software for one is a low-stakes place to test that claim on your own life before betting a business on it.
The pattern Waxman names — "aggregate data, add an LLM" — is also the one most likely to generalize into professional tooling. His fitness app's value isn't the tracking, it's combining nutrition, sleep, weight, and training data that lives in four separate apps and using an LLM to reason across the combined picture, something no single vendor's app does. That's the same shape YC's own Fall 2026 requests for startups describe as a "personal runtime" gap — a category multiple commenters flagged directly in the thread as still unsolved.
Getting started
If you want to try this yourself, the fastest on-ramp mirrors Waxman's stack almost exactly:
npx create-next-app@latest my-personal-app --typescript --tailwind
npx shadcn@latest init
Wire up Postgres on Neon with Drizzle for the schema, add Better Auth if more than one person in your household needs to log in, and point Claude Code at the repo with a plain-language description of what the PDF, spreadsheet, or scattered app data should become. The lower-friction alternatives — Claude Artifacts, Replit, or Lovable — skip the terminal setup entirely if you'd rather not manage a database and deployment pipeline yourself.
Related reading
- Sam Altman’s AI family podcast backlash: tool or substitute?
- Loop engineering with Claude Code: a complete guide
- How to build a website with Claude Code, zero coding required
- Build your first MCP server, step by step
- How to build your first agent skill, step by step
- "Programming is low-intelligence work" — the kache X debate, explained
- YC's requests for startups, Fall 2026
- Clay Shirky, "Situated Software" (2004)
- Alex Waxman's original essay
Costs, stack choices, and quotes reflect Alex Waxman's essay and the accompanying Hacker News thread as of August 1, 2026. Pricing for Claude Code, Vercel, and Neon changes over time — check current vendor pricing before budgeting your own project.