AI Cloud Management: MCP Server and LLM Cost Governance
Two sides of AI and the cloud: letting an assistant work with your estate safely, and keeping the assistant bill in check.
Your cloud, from an AI assistant
ZopNight runs a Model Context Protocol server, so Claude Desktop, Claude Code, Cursor and Codex can work with the same data you see in the app: resources, schedules, costs, recommendations, teams, budgets, audit logs, dashboards, autoscaler policies and event readiness plans. Ask which resources in an account cost the most this month, or which open recommendations touch a given team, and the assistant answers from live data.
Access is personal-access-token based, and an organisation can switch MCP off entirely.
Writes are opt-in and tiered
Every organisation starts read-only. An admin can raise the write tier step by step: metadata-only changes, then reversible actions, then irreversible ones, which additionally require a token bound to the organisation. A global switch keeps the write surface inert until it is turned on. The assistant is only ever shown the tools its caller is allowed to use, and some things are never exposed at any tier: deleting the organisation or cloud accounts, reading credentials, and assigning org-level roles. Role and member management arrive at tier 1, and bulk start and stop at tier 3. Every tool call, including reads, lands in the audit log.
Your AI spend, governed
The AI Gateway gives a cost handle on the LLM calls your own applications make:
- Providers. Connect OpenAI, Anthropic, OpenRouter or AWS Bedrock; a curated cheap and strong model set is registered on connect, and you can add any other model the provider offers.
- Virtual keys. Issue per-team keys with a hard USD budget and a list of allowed models. Applications call one endpoint with one key and switch provider by changing the model name. Keys can be rotated or revoked.
- Routing. A complexity router sends trivial prompts to the cheap tier and harder ones to the strong tier, and you can dry-run a prompt to see where it would go.
- Spend. Month-to-date spend by provider, model, tier and team sits in Cost Reports next to cloud spend, and an AI budget can be set per provider on the Budgets page.
Requests and responses pass through the ZopNight-hosted gateway. The complexity router reads each prompt to choose the cheap or strong tier.
Managed AI services on the cloud side
Model hosting billed by your cloud is covered by ordinary recommendations, for example Bedrock Provisioned Throughput Idle and Idle SageMaker Endpoint, No Invocations. See SageMaker cost.
Further reading: the MCP server, LLM cost governance and virtual keys.
Questions we get a lot.
Can an AI assistant stop my servers through MCP?
Only if an admin has raised the organisation to a write tier that allows it. Every organisation starts read-only, and the write surface also sits behind a global switch.
Does the AI Gateway see my prompts?
Yes. Requests and responses pass through the ZopNight-hosted gateway endpoint, and the complexity router reads each prompt to choose the cheap or strong tier.