Skip to content

Thank you, OSMC 2026

OSMC 2026 wrapped up in Nuremberg today. I went to OSMC for many years, but this was my first time back in several, and the 20 year anniversary was a good one to return for.

A lot was familiar. I met a number of old friends this week, and some new people as well. The new hotel was a great upgrade, though I did miss Checkpoint Jenny. Thank you to everyone who came to my talk, asked questions, or stopped me in the hallway afterwards, and to the NETWAYS team for a well-run conference.

The slides

The deck from my talk is up:

Most of the talk was a walk through the last two years: the new checks, Linux, the transports, the security work. The last part was new, so here is a short summary for those who were not in the room.

What the agent knows

After Nagios World I wrote that context is the piece classic monitoring has always been missing, and that the agent is a natural place to collect it. At the time that was just an idea. It is now four things an agent can tell you about its own host:

  • Host facts: what the host is. An opt-in inventory of the OS, hardware, network interfaces, volumes, installed software, services and more. Already released, since 0.24.0.
  • Metrics: how much. The snapshot of CPU, memory, disk and network the agent has had for years, now also with pressure metrics (is anything actually waiting?) and how much a disk has grown over the last hour and day.
  • Diagnostics: what’s going on right now. Small tables rather than numbers: the top processes by CPU and memory, services that are not where they should be, recent service crashes and event log errors, and the places that usually fill a disk. Off by default, and never a command line.
  • Failed checks: what it looked like when it failed. The agent keeps the last 50 checks that did not come back OK, each with the metrics and diagnostics from that moment. By the time a human logs in the spike is gone. The snapshot is not.

All of it lives on the box and needs no server. You can read it over REST, in nscp test, with check_nsclient or in the web UI.

A first guess from Fleet

The number that matters is the time from error to recovery, and most of that time is not the check interval or the fix. It is a human working out what is actually wrong.

So an enrolled agent can report a failed check to NSClient Fleet together with its context, and Fleet can hand the check result, the host facts, the snapshot and the diagnostics to a model and ask for a first guess. What you got paged with is this:

CRITICAL: 1m: 87%

What Fleet says about it is this:

$ check_nsclient fleet explain --host faults-osl-14 --check check_cpu "CRITICAL: 1m: 87%"

faults-osl-14: The 1-minute CPU load [...] is critical because
acme-indexer.exe is holding the host's only CPU core at 100% [...]
The evidence [...] does not show why the indexer is consuming the
core or whether that behaviour is expected.

That is a real answer to a fault I planted on a test host, shortened to fit. It names the process, it knows the host has one core, and it says what it does not know.

A few things I care about here:

  • It is off by default. Nothing leaves the Fleet server until you configure a provider.
  • You pick the model. Claude, anything OpenAI-compatible, or a local Ollama on your own hardware, in which case the evidence never leaves your network.
  • The bill has a ceiling. One explanation per problem, not per alert, under a daily budget. With a small hosted model it came to a fraction of a cent each in my tests.
  • Your monitoring stays the same. The checks, thresholds and notifications are untouched. This sits at the very end and explains the failure you already got.

And if you do not like AI, the same context is just as useful to a human.

Out on Monday

I had planned to stand on stage and say that all of this is out now. It is not, quite. Host facts and the metrics snapshot are released, but the new metrics, the diagnostics, the failed check log and the Fleet side are coming on Monday.

Thanks again, everyone. I hope to be back next year.

// Michael Medin