openITCOCKPIT

openITCOCKPIT MCP Server

An MCP server that puts an openITCOCKPIT monitoring instance in front of an LLM client: what is broken and why, what someone already handles, what changed - and, off by default, the tools that act on it.

[!CAUTION]

Use at your own risk. This server lets a language model read and - with write

tools enabled - change your monitoring configuration. A model can misunderstand a

request or pick the wrong call. Review what it proposes before you approve it, and

start with read-only access. The software is provided "as is", without warranty

or liability of any kind; see the MIT License.

• Requires openITCOCKPIT 5.6 or newer (compatibility).
• 42 tools, 20 read-only and 22 that change something.
• Write tools are off by default and are not registered until you enable them.
• Names, never IDs. Tools take hostnames, template names and container paths; the server resolves them.
• Scope-checked writes. References are validated against the target container before anything is sent, which openITCOCKPIT's own API does not do.

Quickstart

cp .env.example .env          # fill in the two secrets, see Configuration

docker compose up --build

Point your client at http://localhost:8000/mcp with the bearer token from your .env. Compose reads that same file for the published port, so setting OITC_PORT there moves both sides at once.

Three settings matter to start: MCP_AUTH_TOKEN (what clients present to this server), OITC_BASEURL and OITC_APIKEY (what this server presents to openITCOCKPIT). Everything else has a default - all of it.

What you can ask it

Every tool by name, with its parameters and what it costs in context: [docs/tools.md](docs/tools.md), generated from the code. What a result looks like and what happens when you write: [docs/using-the-tools.md](docs/using-the-tools.md).

Limiting an instance

Running a monitoring instance is not one job, so a surface small enough to be easy would be too small to be useful. Instead the server keeps every tool the work needs and a toolset hands one agent its share: around ten tools, complete for that job and nothing beyond it. A tool that was never registered cannot be called at all.

OITC_TOOLSETS=health          # find things and see how they are doing

OITC_TOOLSETS=operations # acknowledge, maintenance windows, check now OITC_TOOLSETS=lifecycle # take things out of the monitoring, or delete them

The shipped sets are health, operations, lifecycle, patch, reporting, catalog, onboarding, config and provisioning. Sets are defined in a file you can replace: docs/toolsets.md.

Measuring a model

oitc-mcp-eval asks operator questions against your instance and checks whether a model answers them from the tools instead of inventing. It keeps every run in SQLite, so models and changes stay comparable.

oitc-mcp-eval --model your-model --samples 3     # against the configured instance

oitc-mcp-eval --compare # newest run of every model ./scripts/eval-throwaway.sh --model your-model # instance created and removed for you

docs/evals.md.

Documentation

Security in one paragraph

The server holds one openITCOCKPIT credential and never sends it to a client. Clients authenticate with their own token; in delegated mode each request carries a short-lived user token instead, and the server acts as that user with their permissions. Write tools stay unregistered until you enable them. The details, including what to give the openITCOCKPIT user: docs/security.md.

License

MIT.