Use IONOS CLOUD Resources in Agentic Workflows

Who this is for: Platform engineers and developers building automated pipelines, CI/CD workflows, or AI agent systems that need real-time access to IONOS CLOUD infrastructure data.

AI agents can use the MCP server as part of multi-step agentic workflows. The read-only tools are well suited for discovery, validation, and reporting tasks.

Infrastructure discovery

An agent can enumerate your entire IONOS CLOUD environment as part of a larger task:

Scan all my data centers, list every server and its attached volumes, and produce a CSV summary.

This uses list_datacenters, list_servers, and list_server_volumes in sequence.

Pre-deployment validation

Before running Terraform or another provisioning tool, use the MCP server to validate the target environment:

Check whether the LAN internal-net exists in data center prod-fra and confirm there are no conflicting firewall rules for port 443.

Cost reporting

Use Billing and Compute tools together for cost attribution reports:

List all servers across all data centers, group them by location, and pull the usage data for each. Summarise the estimated monthly cost per location.

Cloud DNS audit

List all DNS zones and their records. Flag any zones without SOA records or with more than 500 records.

Certificate expiry checks

List all certificates in Certificate Manager and flag any that expire within 30 days.

Unattended compliance audits (scheduled agent)

Because every tool is read-only, the MCP server is safe to deploy inside an unattended agent loop running on a CI scheduler, cron, or a long-running agent worker. The agent cannot break anything; the worst case is a noisy alert.

A common pattern: every night the agent runs a four-step loop without a human in the chat.

1. Enumerate the resources to audit (for example, list_object_storage_buckets across every region).
2. Run the relevant inspection tools on each (for Object Storage: the six security-posture probes such as get_object_storage_bucket_public_access_block and get_object_storage_bucket_encryption).
3. Compare the result with last night's snapshot stored in a Git repo or an object store.
4. If anything regressed (new public access, encryption disabled, lifecycle removed), open a ticket in your issue tracker with the affected resource, the regression, and a diff against the prior known-good state.

The same agent can run on Claude Code with a project-scoped .mcp.json, on a custom runner built with the OpenAI Agents SDK or LangGraph, or on any other agent framework that speaks MCP. The server is identical in all cases; only the orchestration framework changes.

Rate limits on unattended agents: A nightly audit over a large account can issue hundreds of sequential API calls. If you hit HTTP 429 responses, add short pauses between tool calls and reduce per-step parallelism. For more information, see Troubleshooting.

Composing multiple MCP servers

The MCP server runs alongside other MCP servers in the same AI client. The AI client merges all tool catalogs, so a single prompt can draw on multiple data sources.

With the IONOS CLOUD Documentation MCP

The IONOS CLOUD documentation site exposes its own MCP server at https://docs.ionos.com/cloud/~gitbook/mcp: a GitBook-powered server with documentation search and retrieval tools. Adding it alongside the infrastructure MCP gives the AI assistant both live data and product knowledge in the same session. Both servers are completely independent; you can use either one without the other.

Context window note: Adding a second MCP server increases your total tool catalog. If you use Windsurf (100-tool limit) or another client with a tool cap, set IONOS_MCP_LOAD_MODE=lazy on the IONOS server so only the always-on tools and loader sentinels are registered up front. For more information, see Selective Tool Loading.

{
  "mcpServers": {
    "ionoscloud-docs": {
      "url": "https://docs.ionos.com/cloud/~gitbook/mcp"
    },
    "ionoscloud": {
      "command": "/usr/local/bin/ionoscloud-mcp",
      "env": {
        "IONOS_TOKEN": "your-api-token"
      }
    }
  }
}

Example prompts that use both servers together:

Look up how FOCUS v1.3 billing works in the IONOS CLOUD docs, then pull my actual billing data for this month and produce a FOCUS-compliant summary.

According to the IONOS CLOUD documentation, what is the recommended token scope for Object Storage? Then check my actual account configuration and report whether it matches.

Read the IONOS CLOUD Certificate Manager documentation, then list all my certificates and flag any that expire within the timeframes the docs recommend monitoring.

What changed in the latest IONOS CLOUD release? Then check whether any of those products are in use in my account.

With other tools

To build a workflow that reads IONOS infrastructure data and then commits a report to GitHub, configure both servers in the same mcpServers block:

{
  "mcpServers": {
    "ionoscloud": {
      "command": "/usr/local/bin/ionoscloud-mcp",
      "env": {
        "IONOS_TOKEN": "your-api-token"
      }
    },
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "your-github-token"
      }
    }
  }
}

You can then ask: "List all servers in my production data center and create a GitHub issue summarising any that have no attached volumes."

Combining with other tools

The MCP server integrates naturally with other tools your AI client has access to. Combine it with a file write tool to export infrastructure data to a spreadsheet, or with a notification tool to send alerts when checks fail.

Keep the default eager mode for agent workflows on clients without a tool cap. The MCP server registers all tools at startup by default, which is the correct setting for unattended agents on Claude Code, Claude Desktop, Cursor, VS Code, and similar. Do not set IONOS_MCP_LOAD_MODE=lazy on these clients: in lazy mode, Compute Engine and Object Storage tools only register after an explicit loader-sentinel call, and an agent running a multi-step task will not know to issue that call. Silent failure looks like this: the agent says "I don't have a tool to list your servers" instead of returning an error. To defend against an inherited IONOS_MCP_LOAD_MODE env value from the host shell of a CI runner, pin "IONOS_MCP_LOAD_MODE": "eager" explicitly in the agent's MCP config.

Windsurf is the exception. Windsurf's 100-tool cap takes precedence and lazy mode is the only option. On Windsurf, expect to prompt the agent explicitly with "Load the Compute tools" (or "Load the Object Storage tools") before any step that needs those tools — the agent loop must include that priming step.