Ask Claude about your servers

The Claude connector is part of Pro and Team.

Cloud Viewer provides an MCP server. Add it to Claude, and Claude can read one team’s Cloud Viewer data to answer questions about your infrastructure: inventory, metrics with their alert thresholds, active and past alerts, system logs, job reports, uptime checks and costs. Access is read-only, and the analysis runs on your own Claude subscription.

Set it up

In Claude (web, desktop or mobile), go to Settings → Connectors → Add custom connector.
Enter the server URL https://api.cloudviewer.app/mcp.
Claude opens a Cloud Viewer sign-in. Sign in, choose the team under Grant access to team, and select Allow read-only access.

Any member of a team on Pro or Team can connect it. It does not need an admin.

Other MCP clients that support remote servers with OAuth sign-in connect with the same URL. In Claude Code:

claude mcp add --transport http cloudviewer https://api.cloudviewer.app/mcp

Then run /mcp in Claude Code to sign in and choose the team.

What you can ask

For example: “Are any alerts firing?”, “Diagnose the disk alert on prod-db-1”, “Which of my endpoints is unreachable, and since when?”, “Which server has the address 203.0.113.7?” or “Why did my costs jump this month?”. Claude fetches the projects, charts, log lines, events and cost trends it needs at the time you ask.

Prompt templates

Where your client lists the connector’s prompts, two are included:

diagnose_alert takes an optional alert title or resource name, as shown in the app, and otherwise uses the most severe firing alert. It asks Claude to read the active alerts, the alert history, the resource’s metrics around the firing window, its log lines and the recent changes, then to name the most likely cause, the evidence and the next check or fix.
explain_cost_spike asks Claude to compare month-to-date, forecast and last month, find when the change started in the 30-day trend, find the resources responsible, and look for events around that point.

Tools

Tool	What it returns
`list_projects`	the team’s connected Hetzner projects, with health and resource counts
`get_project_snapshot`	every resource in one project: status, health, specs, labels and dependencies
`get_resource`	one resource from a project: specs, status, labels, location, dependencies
`list_events`	recent provider actions in a project, such as create, delete, attach or rebuild (30 days)
`find_resource_by_ip`	which resources have or contain an IP address
`get_dedicated_servers`	the team’s dedicated (Robot) servers, with agent status and contract prices
`get_metrics`	a resource’s charts with the alert thresholds: CPU, memory, disk, load and network for servers; requests, connections and bandwidth for load balancers
`get_server_metrics`	the agent’s memory, disk and load charts for a dedicated server
`get_costs`	month-to-date and forecast, the daily trend, the top resources, or costs by label
`get_active_alerts`	alerts firing now, most severe first, muted ones marked
`get_alert_history`	alert episodes of the last 7 days, resolved ones included
`get_agent_logs`	recent warning-level system log lines (72 hours), for one server or the whole team
`get_jobs`	a server’s job reports: last run, last success, exit code, staleness
`get_checks`	uptime checks: status, since when, uptime, last error class, certificate days left

Limits

Read-only. Every tool reads Cloud Viewer’s own data. None of them can change a server, a project or a setting.
No Hetzner token. The connector never holds your Hetzner token and cannot call the Hetzner API.
One team: the one you chose when you signed in. Your plan’s limits apply as in the app, for example how far back metrics go.
60 tool calls a minute. Above that, a call returns an error that asks Claude to pause.
64 KiB per result. A longer result is cut off, with a note to narrow the query.
Pro and Team only. Every call checks the team’s current plan, so the connector stops working when a team leaves Pro or Team.

Revoke access

Remove the connector in Claude.

Hand off to a coding agentHow an agent uses the connector and a brief to find the cause and propose a fix.