zopnight

MCP server to query and govern your AWS, GCP & Azure cloud estate (and AI runtimes like Bedrock & Vertex AI) from Claude Code or Cursor surfaces waste, cost, ownership, and 400+ audit findings across 200+ resource types.

View on zopnight.com

Documentation

ZopNight provides a Model Context Protocol (MCP) server that lets AI assistants query your cloud infrastructure data directly. You can ask your AI tool about resources, costs, schedules, recommendations, and more — all through natural language.

What is MCP?

MCP (Model Context Protocol) is an open standard that allows AI tools to connect to external data sources. ZopNight's MCP server uses Streamable HTTP transport — your AI tool sends standard HTTP POST requests to the server. It is not WebSocket based, and there is no STDIO option: this is a remote server, not a local process.

Prerequisites

Before connecting, you need:

  • A ZopNight account with at least one organisation
  • Admin or Editor role in the organisation
  • MCP Server enabled for your organisation
  • A Personal Access Token (PAT)

Follow the How to Connect guide to enable MCP and create your token.

How It Works

  1. Your AI tool sends a POST request to the MCP server URL with your PAT token
  2. The gateway validates the token and identifies your email
  3. The AI tool calls list_organisations to discover your organisations
  4. The server checks that MCP is enabled for the requested organisation
  5. The server verifies you are a member of that organisation
  6. The gateway checks the call against your live role — and, if the tool changes something, against the organisation's Write access setting
  7. Results are returned as JSON to the AI tool

Every check happens at the gateway, on the way in, against the same policy table the UI's own calls resolve against. MCP is not a second way in with rules of its own, and because your role is read per request, revoking access in the app takes effect on the agent's next tool call.

Protocol Details

The MCP server speaks JSON-RPC 2.0 over HTTP POST. It supports initialize, tools/list, tools/call and notifications/*, and it answers ping on the protocol revisions in use today. Anything else answers -32601 method not found, which is correct rather than missing — a server must not advertise a capability it has not implemented. One method moves between revisions: the 2026-07-28 revision removes ping and adds server/discover, so a client that negotiates that revision gets -32601 for ping.

Your client picks the response format with its Accept header. application/json (or no Accept) returns one buffered JSON response, which is what most clients do. A client that sends text/event-stream gets the reply as Server-Sent Events instead, which is what lets a long-running tool report progress while it works. Both are Streamable HTTP; neither is WebSocket.

Connect Your AI Tool

After completing the setup, connect your preferred AI tool:

  • Connect to Cursor
  • Connect to Claude Code
  • Connect to Codex
  • Connect to Windsurf
  • Connect to Antigravity

What Can You Do?

The MCP server exposes read tools covering resources, costs, schedules, recommendations, cloud accounts, teams, budgets, metrics, billing, and audit logs, and a set of write tools that change things. See Available Tools for what each class covers, and ask your AI tool to list its ZopNight tools for the authoritative list.

Write access is one setting per organisation

How much an agent may change is set per organisation, via Write access in Settings → Organisation, with four cumulative levels from read-only through to irreversible changes. An agent can never exceed the permissions of the person whose token it holds.

Before you raise that setting, read the disclosure in How to Connect: enabling MCP write means any content in your cloud environment becomes a potential instruction to your agent.

Having issues? See the Troubleshooting guide for common problems and solutions.