Everything your product depends on, watched from outside.
Including the parts that fail quietly: a lost tool, a fallback model, worse answers.
What it watches
Sites and APIs
Up from several regions, status, speed, the words on the page, JSON fields.
MCP servers and gateways
The handshake, every tool listed and called, changes noticed, time to first token.
Agents
Real questions on a schedule and after deploys, the answers judged.
Jobs and certificates
Heartbeats for scheduled jobs; certificate, domain and DNS changes.
How a failure is confirmed
One failed check is not an outage. When a run fails, two other regions try again at once. It is Down only when two of the three fail.
Flapping folds into one alert: a monitor that flips more than 4 times in an hour sends one message.
How it behaves
- Up and right. Every AI check asks two questions: did it answer, and was the answer right?
- Calm alerts. A second region confirms before anyone is alerted, and one cause makes one incident.
- Explain everything. Every alert says what failed, where, since when, and what changed just before.
See it on your own product
Paste a URL on the home page. No account needed for the first look.
Planned, not in the first release: browser journeys, private probes in your network, imports and a GitHub app.