ZopNight MCP Server — Your Cloud Costs, In Your AI Assistant
What can an AI assistant do with ZopNight?
ZopNight's MCP server exposes 119 tools over a live estate: costs, budgets, anomalies, recommendations, topology, Kubernetes logs, deploys and schedules. Writes pass a three-layer permission model with 4 org-level tiers from off through irreversible. Every call is authorized exactly like the equivalent UI action and audit-logged.
Tools by category
Ask your AI assistant about your cloud spend, and let it act safely. ZopNight ships a native MCP (Model Context Protocol) server exposing 119 tools over your live estate: costs and trends, budgets, anomalies, recommendations, resource topology, Kubernetes logs and events, deploys and schedules. Reads are instant; writes are governed by a three-layer permission model your org controls.
Every tool call authorized like the UI action
Point any MCP-capable client (Claude Code, Cursor, Codex, Windsurf, Claude Desktop) at the ZopNight endpoint with a personal access token. Every tool call is authorized exactly like the equivalent UI action: your user RBAC applies, the org-level write tier (off, metadata-only, reversible, irreversible) caps what any assistant can do, and every mutating tool supports idempotency keys so retries are safe. All MCP activity is audit-logged with its source, alongside UI and API actions.
A personal access token and one client config block
- Create a personal access token: Settings > Developer > New token (choose scopes and expiry; it is shown once).
- Add the server to your client config: {“mcpServers”:{“zopnight”:{“url”:“https://<your-gateway>/mcp-server”,“headers”:{“Authorization”:“Bearer zn_pat_<token>”}}}}. Or: claude mcp add zopnight —transport http —url … —header “Authorization: Bearer …”.
- Ask your assistant for a cost overview to verify.
Read tools, mutating tools, and the three write tiers
119 tools: 85 read, 34 mutating across three write tiers. Cost analytics, budgets, anomalies, recommendations, topology, K8s observability, deploy management, schedule control, integrations and audit logs. Idempotent writes, progress streaming for long operations, full audit trail with source attribution.
MCP server: common questions
Can an assistant stop my instances as soon as I connect it?
No. MCP is off until an org admin turns it on, and the write tier then defaults to read-only. Stopping a resource sits at tier 3, irreversible, which an admin has to select explicitly and which additionally requires token binding.
Does the MCP server bypass my RBAC?
No. Every tool call is authorised exactly like the equivalent action in the interface, so the calling user's own role still applies. The org-level write tier is a ceiling on top of that, never a way around it.
What happens if a tool call is retried after a timeout?
Nothing happens twice. Every mutating tool supports idempotency keys, so a repeated call will not stop the same resource again. All MCP activity is audit-logged with its source alongside interface and API actions.