Real-time monitoring dashboard for Proxmox VE, PBS, Docker, Kubernetes, TrueNAS and vSphere. Self-hosted, with alerts for failed backups, full disks, unhealthy containers and offline nodes.
What happened overnight on my Proxmox and Docker fleet?
You wake up, a service is slow, and you want to know what changed while you
slept. An AI agent with shell access can read what each host still holds when
it asks (pvesh, docker ps, journalctl). Pulse kept watching all night,
so the agent can also ask it what changed across the whole fleet, and when.
This walkthrough connects Claude Code, Codex or OpenCode to Pulse with a read-only token, asks that question, and shows what the agent gets back and what it can't.
What you need
Everything here is free with Pulse Community.
- Pulse v6.7.0 or later. Earlier releases don't have
list_recent_changesorget_metric_history. Install it with Docker or the Proxmox LXC installer from the installation guide. - Your systems added to Pulse. Proxmox connects with a read-only API token under Settings → Infrastructure. Docker hosts need the Pulse agent: Settings → Infrastructure → Install on a host.
pulse-mcpon the machine where your agent runs. Install it with the verified installer steps, then open Settings → API Access → Agent integrations. That tab mints amonitoring:readtoken and generates the configuration for Claude Code and OpenCode. Other MCP clients, such as Codex, use the same command and token.
A monitoring:read token can read history and status. It can't change
anything, and it never receives command text or output.
Ask the question
In your agent, ask in plain words:
What happened on my Proxmox and Docker fleet overnight? Use Pulse.
Pulse gives the agent four reads for this. The agent chooses which to call; here is what each returns, using Pulse's demo data.
1. What changed: list_recent_changes
The agent asks for changes since last night, for example
since: 2026-10-28T22:00:00Z. Without since it gets the last 24 hours.
Changes come back newest first: guests and hosts stopping, starting or
rebooting, containers restarting, resources appearing or disappearing, alerts
firing and resolving, and configuration or relationship changes. Here is one
alert, trimmed:
{
"observedAt": "2026-10-29T19:06:22Z",
"resourceId": "agent-b9070eb05b18bb77",
"resourceName": "Field Office Edge 01",
"kind": "alert_fired",
"reason": "Docker host 'Field Office Edge 01' is offline",
"metadata": { "alert_level": "critical", "alert_type": "docker-host-offline" }
}
A busy night can match more than one read returns (up to 200). The response
then says "truncated": true and gives the full count. The agent should
narrow the read instead of paging: by kind (for example
restart,state_transition,alert_fired), by sourceType, or by one
resourceId.
2. How a metric moved: get_metric_history
Once a resource looks interesting, the agent can ask how its metrics moved, for example whether a disk filled gradually or all at once, or whether memory climbed before a restart. Points are averaged into buckets, each with its minimum and maximum:
{
"resourceId": "nebula-1-mock",
"range": "24h",
"metrics": {
"disk": [
{ "timestamp": 1793214000000, "value": 46.32, "min": 46.28, "max": 46.35 },
{ "timestamp": 1793221200000, "value": 46.50, "min": 46.38, "max": 46.62 }
]
}
}
Community keeps seven days of metric history; a longer range returns
license_required.
3. Where things stand now: get_fleet_context and get_resource_context
get_fleet_context is one read across every resource, with finding and
approval counts, so the agent can see what still needs attention.
get_resource_context gives one resource in depth: its identity, the state
you set (maintenance window, intentionally offline), active findings and
recent actions.
What Pulse can't tell you
The answer is only as good as what Pulse saw. Read these before trusting an empty or short answer.
- Pulse polls. Proxmox is read every 10 seconds by default. A VM that stopped and started between two reads is still recorded as a restart, because its uptime reset. A status flap shorter than one read is not seen.
- Reboots inside a guest. A VM that rebooted from inside, without its QEMU process restarting, keeps its Proxmox uptime. Pulse sees that reboot only if the Pulse agent runs inside the guest.
- While Pulse was down. Nothing is recorded while Pulse itself is stopped or updating. When it starts again it compares each resource with what it last recorded. A state that changed in the gap is recorded once, as "resource state changed while Pulse was stopped", at the time Pulse noticed rather than when it happened. Resources that appeared in the gap are recorded as discovered. Removals and reboots during the gap are not recorded.
- Logs and task history. For multi-guest
vzdumpbackup jobs, Pulse reads Proxmox task logs to derive per-guest backup results. This bounded backup collection is not a general task-log archive:list_recent_changesand the other MCP reads above don't return raw Proxmox task logs, system journals or container logs. Your agent needs separate authorised host access to read those directly; a Pulse token doesn't grant it. - Retention. The change history keeps 30 days. Metric history keeps seven days on Community.
- An empty answer is not proof that nothing happened. It means Pulse recorded nothing in that window.
Get changes as they happen
pulse-mcp --emit-notifications also forwards alerts firing and resolving
and status changes as they happen. Each event carries the changeId that
list_recent_changes returns, so the agent can read the history around it.
Not every MCP client shows these notifications; see
What ships to help you connect.
Tell us how it went
If this answered a real question for you, or failed to, a short note in GitHub Discussions helps decide where this goes next. It's never required.