Better Stack

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 monitoring and on-call that replace Better Stack — and know why this is ALMOST.

Uptime checks are a weekend. **On-call routing is not**: a rota, escalation to a second person when the first does not answer, acknowledgement, overrides for holidays, and a delivery channel that wakes somebody at three in the morning. That last part means a phone call or a push notification through a registered application, which is a provider and a cost. If nobody is on call, skip half this document.

STACK
- Node 20+ with Fastify for the control plane
- SQLite through better-sqlite3, WAL mode
- Probes as small agents on cheap machines elsewhere
- Caddy in front
- **On different infrastructure from what it watches.** A monitor on the machine it monitors reports nothing when it matters

THE DATA MODEL
- monitors, checks, incidents — the usual shape, checks append-only with timings split into DNS, connect, TLS and time to first byte
- schedules: id, name, timezone, rotation_kind, handoff_local_time, layers_json
- shifts: id, schedule_id, user_id, starts_at, ends_at, is_override — computed from the rotation and materialised, so 'who is on call right now' is one indexed lookup rather than a calculation at the worst moment
- escalation_policies: id, name, steps_json — each step a target, a channel and a delay
- pages: id, incident_id, step_index, user_id, channel, sent_at, delivered_at, acknowledged_at, error
- acknowledgements, resolutions, and a post-incident timeline assembled from all of it

ON-CALL, PROPERLY
- A rotation with layers: primary and secondary, weekly or daily, with a handoff time in a named zone. Daylight saving must not create an hour with nobody on call, and that is a real test to write
- Overrides for a holiday or an illness, added by anybody, visible to everybody
- Escalation: page the primary; if unacknowledged after N minutes, page the secondary; then the whole team. Every step recorded with its delivery result
- Acknowledgement from the notification itself, needing no login
- A schedule preview showing the next month, and a warning when a gap exists. A rota with a gap is discovered at three in the morning

WAKING SOMEBODY UP
- Email will not do it. The options are a phone call or an SMS through a provider, or a push notification through your own registered mobile application
- The phone call is the reliable one and it costs per minute. Say so in the README, and set which severities may use it
- The delivery path must not depend on the infrastructure being watched, including the mail server
- A test page, on demand, so everybody has proved their own notifications work before they are needed

NOT CRYING WOLF
- Confirm from more than one region before opening an incident, and re-check immediately from elsewhere before declaring anything
- Deduplicate: one bad deploy is one incident, not nine hundred
- Flapping suppressed, recovery always notified, muting always with an end time
- Quiet hours per monitor with a severity that ignores them

LOGS, IF YOU WANT THEM HERE
- Do not build a log store into this. Ship logs to a purpose-built one and link to a filtered query from the incident. Combining the two is what makes this product expensive and it is not where the value is

THE STATUS PAGE
- Rendered to static files on every change and served from somewhere else entirely, so it survives the application being dead

OPERATIONS
- .env: DATABASE_PATH, BASE_URL, SMTP_URL, VOICE_PROVIDER_KEY, PROBE_TOKEN_SALT, SESSION_SECRET, DEADMAN_URL
- Migrations on boot, each once; nightly backup; a dead-man's switch on the scheduler itself
- Health endpoint reporting every probe's last contact and the current on-call

WHAT MATTERS MOST
The rota with no gaps and a delivery path that wakes people. Build those and rehearse a real page before relying on any of it.