Can I self-host the NIS2 walkthrough's MCP server?

What do you have to set for the MCP server?

There is no separate setting for the MCP server, because it is part of the application. Every instance, nisd2.eu and each self-hosted one alike, offers it at /api/mcp, with the same five tools and the same checks on every call.

For the browser sign-in, the instance is its own sign-in service under OAuth 2.1. There is no third-party service and no identity you apply for with a vendor: an assistant registers itself with your instance.

The instance reads all these addresses from NEXT_PUBLIC_APP_URL: the sign-in service's, the endpoint's, and the origin a browser may call it from. Set the variable to the address your people type.

If it is missing, the instance gives https://www.nisd2.eu as its address, and an assistant's sign-in goes there instead of to your instance.

NEXT_PUBLIC_APP_URL=https://isms.example.com

Who can connect an assistant on your own instance?

An assistant works in the NIS 2 walkthrough, and on a self-hosted instance new accounts do not get the walkthrough. It is unlocked for the accounts whose address is in PLATFORM_ADMIN_EMAILS. So these are the people who can connect an assistant; everyone else sees the way to the walkthrough in Settings instead of a key.

Anyone in PLATFORM_ADMIN_EMAILS also gets the /platform-admin area, with which the operator manages the whole instance. List only people who may have that anyway. Their assistant does not reach those functions, because they refuse every call made with a key.

Beyond that, the same applies as on nisd2.eu: a company admin turns on “Allow AI assistants for this company” in Settings, and the switch is off until then. There is no switch for the whole instance, such as an environment variable.

Which assistants reach an instance inside the company network?

Clients that run on a person's computer connect from there. Claude Code, VS Code, Cursor and Gemini CLI therefore also reach an instance that sits only inside the company network.

We have tested Claude Code, with the browser sign-in and with a key. Set up the others from their makers' documentation.

Claude on the web and desktop, ChatGPT and Copilot Studio connect from their vendor's cloud. Anthropic says that Claude calls remote MCP servers from its cloud, not from the person's device. These clients need an instance reachable from the internet over https; we have not tested them against an instance yet.

The MCP specification requires https for all of the sign-in service's addresses. So run the instance behind TLS inside the company network too, as the documentation describes under domains and TLS.

claude mcp add --transport http nisd2 https://isms.example.com/api/mcp

Does it work with no cloud at all?

With a client on your computers and a language model on your own hardware, nothing leaves your network: instance, assistant and model are all with you. This is the route for companies that want to give no data from the walkthrough to an AI vendor.

The client has to speak MCP over HTTP and be able to use a local model. We have not tested such a setup against the walkthrough yet.

Check a local model's entries like any other assistant's: every screen it wrote shows the name its client gives and the day.

What happens when you change AUTH_SECRET?

The instance signs keys and sign-ins with a value it derives from AUTH_SECRET. It does not store the keys themselves.

If you change AUTH_SECRET, every key, every sign-in and every assistant registration ends, and each assistant is connected again. That is also the route if you fear the secret has become known.