1 What you can build right now
With the current setup you can ship real, internet-facing web apps — by yourself, by talking to Claude.
Web apps & PWAs
Dashboards, trackers, tools, small SaaS — anything that runs in a browser. They can be installable on a phone home screen (PWA).
Apps with a backend
A small server, an API, saved data, scheduled jobs — not just static pages.
AI features inside the app
Let your app call Claude (or other models) — explanations, summaries, chat, generation — billed per use.
Automatic deployment
Keep a change → it builds and goes live on its own, with a real domain and HTTPS.
Many apps, one server
Each app gets its own address app.l.eyri.app. Spinning up a new one is a few minutes.
Self-checking quality
Tests run on a schedule; when something breaks, an AI agent is asked to fix it and opens a proposal for you.
The first app on this platform — MarathonTrainer — was built and deployed exactly this way.
2 Getting started — entirely in the cloud
No installs on your computer. You only need a browser and the two accounts below.
- Open Claude. Go to claude.ai and sign in. This is where you'll do all the building — by chatting.
- Describe your idea. Tell Claude what you want in plain words: "Build a small web app that tracks my reading list, with a page per book." Claude asks clarifying questions, then writes the code in a cloud workspace connected to GitHub.
- Watch it take shape. Claude shows you a preview and explains what it did. Ask for changes the same way you'd ask a person: "make the buttons bigger," "add a search box."
- Send it live. When you're happy, Claude saves the change (a "push" to the main branch). That automatically builds the app and deploys it. A minute or two later it's live at your-app.l.eyri.app with HTTPS.
- Use it from anywhere. Open the address on your phone or laptop. On a phone you can usually Add to Home Screen so it behaves like an app.
- Fixing issues. If something looks wrong, just tell Claude. For bugs the tests catch on their own, an AI agent is asked to investigate and opens a proposed fix (a "PR") for you to approve — you stay in control. Anything that needs a real decision becomes a tracked issue with a one-click way to hand it to a cloud session.
3 What it costs
Three separate things can cost money. You can start with just the first.
| What | Price | When you need it |
|---|---|---|
| Claude Pro the AI that builds your apps |
$20 / mo | From day one — this is your "developer." Next tier: Claude Max ($100 or $200/mo) for much more usage if you build a lot. |
| The server (VPS) where your apps live online |
≈ 830 ₽ / mo (~$9) |
when needed Only once you want apps reachable on the internet 24/7. One server hosts many apps. |
| AI used inside your app e.g. an "explain" button that calls Claude |
pay per use | Only if your app itself calls an AI. You pre-load credit (e.g. $5) and it's drawn down per request. |
"Pay per use" in plain numbers
When your app calls Claude, you pay for the size of each request — but you don't have to think in those units. A practical rule of thumb for a typical short request (like MarathonTrainer's "explain this workout"):
per "explain" request
requests for $5 of credit
requests for $20
Rough estimates for a Claude Sonnet-class model on short prompts; a longer or chattier feature costs more per call. You set a spending limit, so it can never surprise you.
4 How it works under the hood
For the curious or technical reader. You don't need any of this to use the platform.
Stack
- GitHub + Actions — source of truth and the CI/CD engine.
- Docker — each app ships as a self-contained image.
- GHCR — private registry for those images.
- Traefik v3 — reverse proxy; routes by domain, issues free Let's Encrypt certificates automatically.
- One VPS — a small Ubuntu server hosting every app side by side.
CI/CD & quality governance
- Every push runs unit tests; they must pass before an app can deploy.
- Nightly the unit tests run on a schedule to catch drift.
- Friday night end-to-end browser tests run against a fresh build.
- On failure a deduplicated issue is filed and handed to an AI agent (@claude) to fix.
- Fixes arrive as PRs — never auto-merged. You review and approve.
Adding a new app is one command plus a push — the platform creates its folder, certificate and route on first deploy. No manual server work.
5 What it can't do (yet)
Honest boundaries, so you know when you'd need to change the setup.
cloud limits Not buildable purely in the cloud
- Native mobile apps (App Store / Google Play). You can build an installable PWA instead, which covers most needs.
- Desktop / native programs that must be installed on a computer.
- Anything needing special hardware — e.g. training AI models on GPUs.
- Work that needs your local machine — reading files on your PC, USB devices, etc.
setup limits Not possible without changing the platform
- Always-on background processing (heavy queues, real-time streaming at scale).
- Large managed databases or big file storage — the current server is modest; these need adding.
- High traffic — one small VPS suits personal and small projects, not millions of users.
- Strict compliance / isolation needs (separate networks, audited infrastructure).
None of these are dead ends — they're "next step" decisions. When you hit one, that's the moment to add a piece of infrastructure, and the same chat-with-Claude workflow helps you do it.