Ctrlk
Log InSign UpSend Feedback
For the complete documentation index, see llms.txt. This page is also available as Markdown.

Fully Sovereign AI Workflow

This use case shows how to keep an AI-assisted IONOS CLOUD workflow entirely within EU infrastructure: from the prompt you type, through inference, to the API calls the assistant makes on your behalf.

Why a sovereign path

A standard MCP setup splits responsibility across two boundaries:

  • The MCP server runs locally and calls the IONOS CLOUD API directly. That traffic stays inside the EU.

  • The AI client sends your prompts and the JSON returned by MCP tools to its model backend. For Claude Desktop, Cursor, or VS Code Copilot, that backend sits outside the EU.

For workloads where resource names, IP addresses, billing figures, or audit-log entries must not leave EU jurisdiction (regulated industries, public sector, contractual data-residency commitments), replace the non-EU model backend with the IONOS CLOUD AI Model Hub. Inference then runs on open-weight LLMs hosted in IONOS's German data centers, and no component of the round trip crosses the EU boundary.

Architecture

Every endpoint in this diagram is operated by IONOS in Germany. No prompt, tool output, or API request reaches a non-EU AI provider in the request path.

Note: Sovereignty applies to the data plane shown above. AI clients may still emit telemetry, crash reports, or autocomplete suggestions to their vendor backend independently of the configured inference endpoint; disable client telemetry if this matters for your compliance posture.

What you need

  • The ionoscloud-mcp binary. For more information, see Quick Start.

  • An IONOS CLOUD API token scoped to the products you intend to inspect. For more information, see Authentication.

  • Access to the IONOS CLOUD AI Model Hub with a token for its OpenAI-compatible endpoint.

  • An AI client that supports both MCP and a custom OpenAI-compatible base URL. The MCP server is client-agnostic, but the AI client itself must let you redirect inference away from the vendor default. Options that meet both criteria today include Continue, OpenCode, and clients built on the OpenAI Agents SDK. See each client's documentation for the exact provider-override setting.

Choose a model

The AI Model Hub serves several open-weight LLMs that support tool calling, a prerequisite for MCP. Reasonable starting points:

Model
Good for

Mistral Small 24B

General infrastructure Q&A, fast responses

Llama 3.3 70B

Multi-step agentic workflows, larger context

GPT-OSS 120B

The most capable open model, when latency is less critical

For more information, see Models Comparison for the full list, context windows, and tool-calling support.

Configure the client

Point your AI client at two things: the IONOS CLOUD AI Model Hub for inference, and the local MCP binary for tools.

  1. In the client's model/provider settings, set the base URL to the AI Model Hub OpenAI-compatible endpoint and provide your Model Hub API token. For more information, see AI Model Hub API for the exact URL and authentication flow.

  2. In the client's MCP settings, add the ionoscloud server entry exactly as shown in the per-client guide (for example, Continue or OpenCode).

  3. Restart the client so it picks up both the new provider and the MCP server.

Verify the sovereign path

Ask the assistant a question that requires a tool call:

List my IONOS CLOUD data centers and summarise them.

If the response includes your real data centers, both halves are wired correctly: the MCP server reached the IONOS CLOUD API for the list, and the Model Hub model produced the summary. Confirm in the client's logs that requests went to your configured Model Hub base URL rather than the vendor default.

Limitations to be aware of

  • Open-weight models trail frontier models on complex agentic chains. For deep multi-step reasoning across dozens of tool calls, expect to iterate more than you would with a frontier closed model. Start with concrete, narrowly scoped prompts.

  • Not every MCP-capable AI client allows a custom inference endpoint. Vendor-locked clients (Claude Desktop, Cursor) currently route inference to their own backend and cannot be redirected to the AI Model Hub. Choose the client based on this constraint, not just on MCP support.

  • Sovereignty is end-to-end. If you later add a third MCP server that calls a non-EU SaaS, that traffic leaves the perimeter; the IONOS-side guarantees apply only to the IONOS CLOUD MCP server and the AI Model Hub.

Related topics

PreviousFinOps and FOCUS BillingNextTroubleshooting

Last updated

Was this helpful?