Skip to content

Written by a scheduled job, not by you loading this page

Everything is answering.
Here is how I know.

A status page that pings the homepage when you open it tells you the CDN is up, which is the least interesting fact available. The homepage is static and it will keep answering long after the MCP server, the Python box or the model key have stopped working.

So each probe below does the smallest real version of the job it is watching. The MCP one opens a session. The Python one reads a PDF. The guard one tries to fetch the cloud metadata address and passes only if it is refused, because a security control that has quietly stopped working looks exactly like one that is fine.

uppagesthe seven pages a visitor can open
100%
uptime over 43 checks
1431ms
median
every 10 min
probed
upguardthe address guard, checked by trying to defeat it
100%
uptime over 43 checks
322ms
median
every 10 min
probed
upmcpthe MCP server, checked by opening a real session
100%
uptime over 43 checks
1917ms
median
every 10 min
probed
uppythonthe Python service on its own box, checked by reading a real PDF
100%
uptime over 43 checks
952ms
median
every 10 min
probed
upintegrationthe GitHub client, checked against a real repository
100%
uptime over 8 checks
1424ms
median
hourly
probed
upevalsthe golden set, which must still score 100% on precision
97.7%
uptime over 43 checks
772ms
median
every 10 min
probed
2026-09-07 10:28 the eval is no longer at 100% precision

How it runs

An n8n workflow on a VPS calls a protected endpoint every ten minutes, and the whole set once an hour because two of the probes cost money or hit somebody else’s server. Never a platform cron: one place to see what is scheduled beats a schedule that exists only inside a deployment config.

The endpoint answers 503 when a probe fails, so the run goes red in the execution list rather than looking like a success that happened to contain bad news.

The first version of the eval probe read only the first 4,000 characters of the page and then looked for a number that appears at about character 6,000, so it reported the eval as broken when it was fine. A monitor that raises a false alarm is worse than no monitor: the first one gets investigated, the third gets ignored, and the fourth is real.

The oldest check shown is from 2026-09-07 10:28. Uptime here means uptime since then and nothing longer, because there is no data before it and a percentage over a short window is easy to mistake for a promise.

The whole list is 41 requirements taken from 114 job posts, with the gaps shown at the same size as the wins.