Experimental MCP Server Feature Coming to Checkmk

Hello Community!

Today, we launch the Checkmk MCP Server, free for Pro, Ultimate/MT, and Cloud editions. Use your AI tool of choice (including local, self-hosted LLMs) to directly access your live site, and run natural language queries like “What needs my attention right now?” to get instant insight into your monitoring data.

A few practical details on how it handles security and data:

  • Disabled by default: turn it on under Setup → Global settings.

  • Strictly read-only: honors native RBAC and will not write changes to your site. Checkmk can not view or store your queries.

  • Air-gap friendly: pair it with a local model so zero telemetry leaves your firewall.

  • Editions: It’s free for Pro, Ultimate/MT, and Cloud editions.

As an experimental release, we want your feedback to shape the future of our AI capabilities.

Full setup instructions and test site results are in the announcement blog post.

Please note: If you update a site rather than installing from scratch, we recommend running the following command: sudo omd update-apache-config <site>

We’re also hosting a Community Call with a demo and engineering Q&A on September 23 at 3 PM CEST on Zoom. You can get the ics file here (1.0 KB).

Subscribe to receive a reminder about the call 1 day and 1 hour before the event: Checkmk Community Event Reminder

Looking forward to seeing you there!

3 Likes

Tested on 2.5.0p13.pro via Claude Code.

First, credit where it is due: connecting took two minutes, and thanks to dynamic client registration there was no user and no secret to create anywhere. The tool descriptions are unusually careful - they document the traps, not just the parameters - and the thing genuinely saves time: questions that used to mean clicking through several views are now one call. Which is why I would like to see one detail addressed before this goes wider.

MCP server: credentials in cleartext - and what does the usage-data switch send?

I called read_setup_configuration for one host because I wanted its IP address. The response included:

"snmp_community": {
  "auth_protocol": "MD5-96",
  "auth_password": "<cleartext>",
  "privacy_protocol": "CBC-DES",
  "privacy_password": "<cleartext>"
}

This matches the REST API’s long-standing behaviour, so it is not a regression. But the consumer is new: “strictly read-only” protects against modification, not against disclosure, and this client forwards what it reads into the context window of an external language model. An agent also enumerates, where a human looks up one host and stops.

Suggestion: mask secret-bearing fields in the MCP layer by default (snmp_community, management_ipmi_credentials, special agent credentials). Junos does this well — a login class without the secret permission sees /* SECRET-DATA */ and everything else still works. Failing that, document it before the consent screen: a monitoring-only role never reaches this tool, but I only learned that after seeing the passwords.

And one question: what exactly does Send MCP usage data to Checkmk transmit? I found nothing in the werk, the blog post, the docs or the inline help. Above all — are tool results included? The result of read_setup_configuration contains those passwords, so with that box ticked the same call would disclose them twice: to the model and to you. Most likely the answer is “counters only”, and one sentence in the inline help would settle it.

the documented settings path does not match the GUI

Both Werk #22263 and the announcement blog post say:

Global settings > Site management > Enable MCP server

In 2.5.0p13 the settings actually live in their own section:

AI features
  Enable MCP server
  Send MCP usage data to Checkmk

Anyone following the instructions looks in the wrong place first. Worth a one-line fix in the werk, the blog post and the docs — it is the very first step a new user takes.

2 Likes

Hello,

thank you so much for your feedback and trying out the MCP server, it’s really nice to hear that you found it useful so far.

Thank you also for your suggestion about password handling, we are taking up your suggestion and will implement it as soon as possible. I will update here to let you know when the fix is available.

Regarding the usage data:

We have a section in the Knowledge Base article available where we describe in more detail what an opt-in would entail. If you choose to opt in we collect the following:

  • Tool and resource calls (this only means the name of the tool, we aren’t collecting the actual prompt or the responses you receive).
  • Call latency
  • Response size
  • AI client
  • Session ID
  • Protocol version

Let us know if you have any other concerns or questions and once more, thank you so much for your feedback!

Annika

2 Likes

And … done:

1 Like

Thank you to everyone who joined the call!

Here is the presentation, in case you were not able to join us today, or would like to review the content:
The Checkmk MCP Server Community Call.pdf (750.1 KB)

1 Like

The MCP can make existing data accessible, but it cannot provide diagnostic context that was never captured. Capturing relevant diagnostic context and service dependencies should receive the same strategic attention as MCP and AI themselves.

What I am currently missing from the roadmap is a broader vision beyond MCP itself: how does Checkmk intend to create meaningful operational value with MCP and AI when the context required for analysis is often unavailable?

Many checks record states and metrics, but not enough context to explain why a service failed or how that failure relates to other services. When a CPU or memory alert is raised, for example, there is often no snapshot of the processes running at that time. When an HTTP check fails, relevant details such as response headers, redirects, TLS errors, or error information contained in the response body may not be captured.

Without this evidence, MCP and AI can do little more than infer probable explanations instead of investigating actual causes. Like a high-performance engine without fuel, the technology may appear impressive, but it cannot deliver meaningful operational value.

I would therefore like to see a roadmap for enriching high-value checks with diagnostic context, retaining it in a structured form for an appropriate period, modeling relevant service relationships and dependencies, and exposing this context through MCP to both operators and AI systems.

How does Checkmk plan to address this context gap? Without a stronger diagnostic foundation, AI integration risks becoming an impressive showcase rather than a genuinely useful tool for root cause analysis.

2 Likes

Hey Lars,

The feedback from users and our internal tests have so far shown that our AI integration provides already as of today high value and unparalleled insights. So, I would argue even what we already have today is beyond an impressive showcase. People use it in production and are happy!

But you are of course also right :slight_smile:

On the conference this year, we shared openly our plans beyond the MCP: https://youtu.be/s61o8UXRzZc?si=hcSAgHEHZhNPw3iu&t=4462

We have a clear vision which goes far beyond the MCP with many more capabilities, also based on a much richer data collection of Checkmk.

We will share more detailed plans for 2027 hopefully soon. There are many exciting things waiting.

1 Like

Hi Martin

As you know I’m a big fan of this initiative. Even in its current form, it already provides valuable insights.

The conference focused mainly on new capabilities and data sources. What remains unclear is how Checkmk plans to make the data already collected by established agents and existing checks available for diagnosis and analysis.

Beyond these new capabilities, how will Checkmk make its existing agents, checks, and monitoring data AI-ready?

For example, a CPU or memory alert could include the process list from the time of the alert. Data retrieved later through MCP may already reflect a different state. Does Checkmk plan to retain relevant history or capture configurable diagnostic snapshots when alerts occur?

This also requires an AI-ready approach to check design. Whenever a check is added or enhanced, relevant numerical metrics should be exposed as performance data, and meaningful diagnostic context should be included in the service details. This should be a standard part of development and review, not deferred until later.

Finally, are service-to-service dependencies planned for alert handling and root-cause analysis? Host dependencies offer limited value when an application is affected by another service while the underlying hosts remain available.

In summary:

  • Preserve diagnostic data: Retain relevant agent and check data, or capture diagnostic snapshots when alerts occur.
  • Design checks for analysis: Expose relevant metrics as performance data and include diagnostic context in service details from the outset.
  • Support service dependencies: Add service-to-service dependencies for alert handling and root-cause analysis.

I hope the plans for 2027 will also provide some insight into these topics.

Hey Lars,

the agent provides a large set of data, which can be very helpful for analysis. Same goes for the topic of relationship between systems, either inferred (implicit) or explicit.

We have included in 3.0 already a generic way to store explicit host relations, starting with mgmt boards - this is a semantic algorithm.

We might go for different paths as you suggested, but the idea is the same: use the agent to collect the information needed for analysis and which is not stored in services.
And, it is on our plan for 2027

Could you elaborate on how the Checkmk MCP works in an Ultimate Edition multi-tenant distributed environment with several sites?

  1. Does the MCP endpoint on the central site provide data from alle remote sites, or only data from the central site itself?
  2. Does the central MCP endpoint forward requests to the MCP endpoints of remote sites, or does the LLM client need to connect to each site endpoint separately?
  3. How should customers without access to the central site connect? If a customer operates multiple sites, is the intended setup to use one client and one OAuth registration per site, or is there a consolidated approach?

Any clarification would be greatly appreciated.

MCP → REST-API and Livestatus.
Central REST-API and livestatus can get data from remote sites

1 Like

Provider LLM Client → Central MCP Endpoint → API/Livestatus → Data from all sites

For the provider, only the MCP server on the central site needs to be enabled.

For customer access, the MCP server must be enabled on the customer site, and the customer connects their LLM directly to it.

@mschlenker
Please add to the documentation which MCP server needs to be enabled in each case.

Team knowledge is currently moving all MT related documentation around the User Guide anyway, so we will consider your findings.