OneSignal

Promptfree, for everyone, and the only version there is

Copy the prompt for Claude CodeCopy the prompt for CodexCopy the prompt for Cursor

Copy

Build me push notifications that replace OneSignal — and read the obstacle first, because it decides everything.

**Delivery goes through Apple and Google, and there is no way around them.** Web push you can do yourself with VAPID keys and no account at all. Mobile push needs a registered application on each platform, which means developer accounts and a release process. That is the barrier, not the code — and the second half of the work is **token lifecycle management**, which is dull and unavoidable.

STACK
- Node 20+ with Fastify
- SQLite through better-sqlite3, WAL mode
- Web push directly with VAPID; APNs and FCM for mobile if you have the applications
- A worker for sending, separate from the web process
- Caddy in front

WEB PUSH, WHICH YOU CAN HAVE TODAY
- Generate a VAPID key pair once. No account, no fee, no third party
- A service worker on your site subscribes the browser and sends you an endpoint and two keys
- You encrypt the payload and POST it to that endpoint. The browser's own push service delivers it
- That is the whole protocol, and it works on every desktop browser and on Android. On iOS it works only for an installed web application, which is the caveat to write down

THE DATA MODEL
- devices: id, user_id, platform, token, endpoint, keys_json, app_version, os_version, language, timezone, created_at, last_seen_at, invalid_at, invalid_reason
- One row per device, never per user. A person has three, and two of them are dead
- notifications: id, title, body, url, data_json, image_path, kind, created_by, created_at
- targets: id, notification_id, device_id, scheduled_for, status, attempts, provider_message_id, sent_at, delivered_at, clicked_at, error — one row per device, which makes a resume exact and a duplicate impossible
- preferences: user_id, category, is_enabled, quiet_hours_json
- segments: saved rules over user properties and behaviour, evaluated at send time

TOKEN LIFECYCLE, WHICH IS THE UNGLAMOROUS HALF
- Tokens expire, are revoked, and migrate when an application is reinstalled. Sending to a dead token wastes quota and, in volume, harms your standing with the platform
- Every rejection tells you why. A permanent failure marks the device invalid immediately; a transient one retries with backoff. Never keep retrying a token the platform has told you is gone
- Deduplicate: the same user on the same device after a reinstall is one device, matched on the platform's own identifier where available
- Prune devices not seen in a long time, and record why they went
- Without this, a device table becomes mostly rubbish within a year and every send takes ten times longer than it should

SENDING
- The worker claims a batch of targets with a lease, sends, records the provider's response, and continues after a crash exactly where it stopped
- Rate limited to what each platform accepts, with backoff on throttling
- Batched where the platform supports it, one at a time where it does not
- A test send to your own devices required before any broadcast

RESPECTING THE PERSON
- Categories, with a preference per category. Everything opt-in
- Quiet hours in the recipient's own zone, enforced in the send path. A notification at three in the morning is an uninstall
- A frequency cap per person per day, enforced centrally rather than per campaign
- A clear way to turn each category off inside your application, not buried in system settings
- Push is the most intrusive channel there is. Every one of these rules is what keeps the permission you were granted

MEASURING
- Sent, delivered where the platform reports it, opened, and the resulting action
- Delivery is not guaranteed and not always reported. Say so beside the number rather than presenting a send count as a delivery count
- Per notification, per platform, per segment, computed at query time

OPERATIONS
- .env: DATABASE_PATH, BASE_URL, VAPID keys, APNS and FCM credentials, SESSION_SECRET
- Migrations on boot, each once; nightly backup off the machine
- Health endpoint reporting invalid-token rate and queue depth

WHAT MATTERS MOST
Token cleanup and quiet hours. The first keeps sending fast and your standing good; the second keeps the permission you were granted, and losing that permission is the one failure you cannot undo.