Nightwatch
Your agent died at 3am.
Nobody noticed until the invoice.
A dead-man's switch for always-on AI agents. Heartbeats, spend-anomaly alerts, and a kill relay — in your Telegram, not another dashboard you won't open.
What it watches
- Heartbeat — your agent pings a URL each loop. Miss the window → you know in seconds, not days.
- Spend rate — token/API burn suddenly 10× baseline? That's a loop, not throughput. Alert before the bill.
- Stall & drift — alive-but-stuck detection: same state hash for N cycles, no progress, frozen inputs.
- Dead-man escalation — warn → page → (optionally) call your agent's own kill endpoint. You define the ladder.
- Telegram-first — alerts and a /kill relay where you actually are. Email fallback included.
- Any stack — it's an HTTP ping from your loop. No SDK, no framework lock-in, five lines to integrate.
How it works
Built from scars, not slides. We ran an autonomous trading agent against live
exchanges 24/7 for 23 days — real market data, real funded accounts, paper fills by
design, ~1,000,000 monitored cycles, zero silent failures. Not because nothing broke:
sockets dropped, keys rotated, feeds froze. The safety layer caught all of it.
Nightwatch is that layer, extracted and hosted for your agent.
Early access
We're validating demand before we build the hosted version — honestly, that's the same discipline we ran the trading agent with. Early-access pricing will be $19/mo (founding users keep it).
No spam. One email when the beta opens, one if we decide not to build it — we'll tell you either way.