When a server is on fire right now and you don’t want to SSH in and eyeball top and pg_stat_activity for ten minutes, Diagnostic Captures does it for you. Trigger a capture, HostAtlas samples the running processes and active database sessions for a short window, saves the samples, and the AI assistant summarises what it found.

Note

Available on the Business plan and above.

What you see

Under Diagnostic Captures in the sidebar. The index lists past and running captures:

• Server the capture targets.
• Triggered by (user or automation) and reason (free-text note).
• Started at, duration, sample count.
• Status — running, complete, cancelled, or failed.
• AI summary preview if one has been generated.

Click a capture to open its detail view:

• Top processes — sampled every few seconds; ranked by CPU and RAM at capture peak, with average across the window.
• Postgres activity — active queries, waits, blockers, longest-running statements.
• AI summary — a plain-language explanation of what appears to be causing load.

What you can do

• Start a capture from the server detail page’s Diagnose now button, from Diagnostic Captures, or programmatically via the API / MCP.
• Cancel a running capture at any time.
• Explain with AI — passes the collected samples to the assistant and gets a summary you can paste into an incident.
• Attach to an incident — link the capture from an incident’s timeline for the postmortem.

How it works

• Triggering a capture queues a command to the target agent. The agent begins periodic sampling: process list ordered by CPU, memory footprint, and (on hosts with Postgres) a pg_stat_activity snapshot per interval.
• Sampling auto-stops after the configured window (typically one to two minutes) so a runaway capture cannot fill disk.
• Samples upload to HostAtlas as they’re taken, so partial data is available while the capture is still running.
• When complete, the capture is retained per your plan’s retention. The AI summary is generated on demand — you press the button, the assistant reads the samples and returns text.

Tip

The capture is observational — nothing is killed, restarted, or rate-limited. If you want to act on what you see, use a Recipe or a Recovery rule.

Related

• Slow Request Log — the application-level equivalent for “one slow route” instead of “the whole box”.
• Forensic Mode — time-travel view of the same server across a wider window.
• Recipes — take action once you know what’s wrong.