clusterctl mcp serve

Serve the Model Context Protocol over standard input and output

Synopsis

Run an MCP server on standard input and output for a client started on this workstation. It acts as you: your configuration, your ssh agent, and the context it was started in, which it keeps for its whole life.

The agent can resolve node sets, describe nodes, query Slurm and run any command that only reads. It cannot change anything directly. A change is planned first, which resolves the node set once, refuses protected hosts and shows exactly what would be sent; applying the plan puts the same question the confirmation prompt asks to you, through the client, and the agent cannot answer it. Every plan and every attempt to apply one is recorded in mcp/audit.jsonl under the state directory.

With –confirm=approval the question is left to the client’s own approval of the apply_plan call instead, for a client that cannot ask. Only use it with a client that asks you before every call of that tool.

Commands on the nodes cannot be run, and –yes, –force, –dry-run and –nodes are refused: the server never confirms or forces anything itself.

claude mcp add clusterctl – clusterctl mcp serve –context prod

clusterctl mcp serve [flags]

Options

      --confirm string      who confirms a change: elicit or approval (default "elicit")
  -h, --help                help for serve
      --plan-ttl duration   how long a plan may wait to be applied (default 10m0s)

Options inherited from parent commands

      --config strings        configuration file or directory to read, repeatable (default: CLUSTERCTL_CONFIG or the search path)
      --context string        context to act on (default: the current one)
      --dry-run               report what would be done and change nothing
      --fanout int            how many hosts to work on at once, at least 1; caps the service processors and the names asked at once too, which fanout.max and CLUSTERCTL_FANOUT do not (default: from the configuration)
      --force                 allow protected hosts, and nodes the inventory does not know, to be touched
  -n, --nodes stringArray     node set to act on, for example 'exe[1-10],@rack:R02' (default: CLUSTERCTL_NODES)
  -o, --output string         output format: table, wide, json, yaml, nodeset, name, jq= (default "table")
      --progress string       how to show the progress of a command on standard error: auto, tty, counter, plain, none (default: CLUSTERCTL_PROGRESS, else auto, a live tree when standard error is a terminal; plain writes lines for a log)
      --progress-log string   append the progress events of a command to this file, one JSON object per line, created readable by you alone (default: CLUSTERCTL_PROGRESS_LOG; an empty one writes none)
      --set stringArray       override one configuration value as PATH=VALUE, repeatable
  -y, --yes                   answer the confirmation prompts with yes

clusterctl mcp serve