Server has 5 tools with explicit schemas and descriptions. Naming follows verb_noun convention (cluster_status, list_guests, node_status, guest_power, snapshot_guest). Descriptions are present (avg ~100 chars) but lack actionable context for LLM selection. Parameters have types and descriptions, but lack enums for constrained inputs (e.g., guest_type, action, power actions). Output schemas are not documented. Error handling is present but minimal, no recovery guidance or categorization. Security is strong: credentials injected via environment, mutations require exact confirmation strings, audit logging implemented. Composition is clean: each tool has one responsibility. Major gaps: no pagination support, no output schema documentation, no error recovery hints.

• Add JSON Schema definitions for all output types. Document cluster_status return as {nodes: list[{name, status, ...}], cluster: {...}, ha: list[...]}. Document list_guests return as {items: list[{vmid, name, type, node, status, ...}], total: int}.
• Convert constrained string parameters to enums in schemas. guest_type: enum=['qemu', 'lxc'], action: enum=['start', 'shutdown', 'reboot', 'stop'], snapshot_name validation regex should be in schema pattern field.
• Enhance error messages with recovery hints. E.g., 'Node "prod-01" is not allowlisted. Allowed nodes: ' + ', '.join(cfg.nodes). For invalid guest_type, return 'Guest type must be qemu or lxc, got: {guest_type}'.
• Add pagination to list_guests: accept optional limit (1-500, default 50) and offset parameters. Return {items: [...], total: int, offset: int, limit: int} so LLM can iterate if needed.
• Expand tool descriptions with WHEN and WHY. E.g., cluster_status: 'Get cluster-wide health, node list, and HA resource status. Call this first to understand cluster topology before targeting specific nodes.' node_status: 'Get CPU, memory, disk, and uptime for a specific node. Use after cluster_status to drill into a node's details.'