Nightshift Labs
Nightshift Watchdog is the dead-man's switch for AI agents and automations. If a job stops pinging — or a loop will not stop spending — you hear about it before the invoice does.
Free for 3 monitors. No card. Email alerts included.
classifier-loop is still up. The spend is the incident.
Silence
An n8n credential expires. A Vercel cron is paused. A Claude Code session never reaches its Stop hook. Nothing throws, so nothing pages you. Nightshift marks the monitor down when the next ping does not arrive, and again when it recovers.
Runaway
A retry that retries the retry. Pings can carry tokens and dollars. On Pro, a hard cap or a spike against the recent baseline becomes an alert — runs per hour, tokens per hour, or dollars per day.
01
Interval or cron, plus a grace period. You get a unique ping URL.
02
curl, n8n, Make, Zapier, Actions, Vercel, Python, Node, or a Claude Code hook.
03
Missed schedule, explicit failure, or a spend spike. Email, Slack, Discord, or Telegram.
$0/mo
For a side project that should not die quietly.
3 monitors · 1 person
$12/mo
Spend limits, loop detection, and chat alerts.
25 monitors · 1 person
$39/mo
A shared night shift for a small crew.
100 monitors · 10 people
Those tools are excellent dead-man's switches for cron. Nightshift is built for the same silence problem, plus the failure mode that showed up with agents: a loop that keeps succeeding and quietly spends tokens, runs, or dollars. Each ping can carry tokens and cost, and Pro plans alert when that spend leaves its baseline or crosses a hard limit.
You set an interval or a 5-field cron expression, plus a grace period. After a successful ping, the next deadline is the following slot plus grace. If that moment passes with no success ping, the monitor is marked down and we alert. A later success ping marks it recovered and alerts again.
The checker is designed to run every minute. Detection is the grace period plus up to one minute. On Vercel, a per-minute cron requires a plan that allows minute schedules. The Hobby plan only runs cron once a day, which is too coarse — use Vercel Pro or point any external minutely scheduler at /api/cron/check.
No. Webhook URLs and bot tokens are encrypted with AES-256-GCM before they are written to Postgres. The encryption key lives in ENCRYPTION_KEY, not in the database. Ping URLs are stored as a hash for lookup, with the recoverable token encrypted beside it.
Each monitor accepts 60 pings a minute and each IP address accepts 240, by default. Excess calls get HTTP 429. That is enough for a healthy agent loop and tight enough that a leaked URL cannot fill the database unchecked.