Your own AI connections

Your organisation can run Chat on its own AI: its own infrastructure, such as Azure OpenAI, its own hardware or another OpenAI-compatible endpoint, or its own account with a provider such as Anthropic, OpenAI, Mistral AI or OpenRouter. An administrator adds each connection in the hub and decides whether people can pick it in Chat and whether it can be used through the API.

The agreement with the provider is your organisation's own. The provider is not a komplai subprocessor, your organisation pays it directly, and the usage does not count towards the monthly AI spending cap. Decide, Redact's detection, Index and search in Documents always use komplai's validated models, so connections do not apply to them.

Add a connection

You need to be an administrator of the organisation. The page is there when your organisation has Chat or Documents.

Sign in and open Organisation › AI and processing › Connections (/settings/ai). The page is called AI connections.
Under Your connections, open Add a connection.
Choose the Type. Your account with a provider takes a Provider: Anthropic, OpenAI, Mistral AI, OpenRouter, or Another provider (OpenAI-compatible). Your own infrastructure is an OpenAI-compatible endpoint, for example Azure OpenAI or your own hardware.
Enter a Name. People see it in Chat, for example Azure Sweden.
For your own infrastructure or another provider, enter the API address, Where is data processed? and Does data stay in the EU/EEA? (Yes, No or Don't know). For another provider, also enter the Provider's name. The address must use https and be reachable from the internet.
Enter the Model as the provider writes it. On Azure, it is the deployment's name. On OpenRouter, openrouter/free picks among OpenRouter's free models, or you can name one model.
Paste the API key. It is stored encrypted, and afterwards only its last four characters are shown.
Choose Can be chosen in Chat and Can be used through the API. A new connection starts with Chat ticked and the API not ticked.
Select Add connection. The page checks the connection straight away, without sending a message, and says whether the provider accepted the key and lists the model.

What people see about a connection

Every connection states where it processes data, who runs it and whether data leaves the EU/EEA, on the page and next to the choice in Chat:

Provider	Processed in	Data leaves the EU/EEA
Anthropic	The United States	Yes
OpenAI	The United States	Yes
Mistral AI	France	No
OpenRouter	The United States; OpenRouter forwards each request to the model's provider, which may be in another country	Yes
Your own infrastructure or another provider	Where your administrator states	As your administrator states

For your own infrastructure and another provider, these are your administrator's statements, and the hub shows them as yours.

Free models. OpenRouter's models that end in :free, and the openrouter/free router, carry a notice next to the choice: the provider may store and use what is sent. Use them for testing, synthetic data and ideas.
Routers. A router such as openrouter/free picks a model for each request. The model that answered is shown in Chat and in the usage log, and the API returns it in model.

Check, change or delete a connection

Check the connection tests the key and the model without sending a message. A connection that does not answer within 15 seconds is reported as not answering.
Edit changes any field. Leave API key empty to keep the key; if you change provider, enter a new key.
Delete connection deletes the key too, and the connection can no longer be chosen in Chat or used through the API. A connection that drafting uses cannot be deleted until you choose other processing for drafting.

The same page has Drafting in Documents, templates and Word: where text is sent when someone asks for a draft or a suggestion. Choose Local infrastructure or one of your connections and select Save choice.

How keys and addresses are protected

An administrator types the key once. It is stored encrypted and never sent back to a browser. komplai's staff never receive or handle it.
komplai's operators see your connections read-only, with the host of your own address, but never the key, its last characters or the full address.
The address is checked when it is saved and again before every call. It must use https, may not point to an internal or private network address and may not redirect.
Creating, changing and deleting a connection, and changing the drafting choice, are written to the organisation's Audit log, without the key.
Only the organisation's administrators can open the page. komplai's read-only support access cannot.

Use a connection through the API

A call can run on a connection when:

it uses a personal API key (kp_live_...) with chat:complete. An organisation integration key (kp_org_...) is refused with integration_not_permitted;
the person behind the key has access to Chat, and your organisation's Chat processing is approved;
the connection has Can be used through the API ticked. The connection then shows Value for processing in the API, such as connection:6d3f9a2e-1b4c-4e7d-8a5f-0c2b9e4d7a31.

The choices under AI and processing › Processing cover komplai's own processing; a connection follows its own two switches.

Copy the connection's value from the Connections page.

List its model with GET /v1/models. The list has one entry: the model or router the connection names.

Shell

curl "https://api.komplai.dk/v1/models?processing=connection:6d3f9a2e-1b4c-4e7d-8a5f-0c2b9e4d7a31" \
  -H "Authorization: Bearer $KOMPLAI_TOKEN"

Response

{
  "object": "list",
  "data": [
    {
      "id": "openrouter/free",
      "object": "model",
      "owned_by": "openrouter",
      "name": "OpenRouter, gratis modeller",
      "processing": "connection:6d3f9a2e-1b4c-4e7d-8a5f-0c2b9e4d7a31"
    }
  ]
}

Send POST /v1/chat/completions with the value in processing. Leave model out, or send the model the connection names.

Shell

curl https://api.komplai.dk/v1/chat/completions \
  -H "Authorization: Bearer $KOMPLAI_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"processing": "connection:6d3f9a2e-1b4c-4e7d-8a5f-0c2b9e4d7a31", "messages": [{"role": "user", "content": "Hvor lang tid har jeg til at klage over en afgørelse om boligstøtte?"}], "max_tokens": 256}'

Read the answer. model is the model that answered, which a router chooses per request; requested_model is the model or router the connection names, and processing repeats the connection.

Part of the response

{
  "object": "chat.completion",
  "model": "meta-llama/llama-3.3-70b-instruct:free",
  "requested_model": "openrouter/free",
  "processing": "connection:6d3f9a2e-1b4c-4e7d-8a5f-0c2b9e4d7a31",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "Du har normalt 4 uger til at klage, regnet fra den dag du fik afgørelsen. Fristen står i afgørelsens klagevejledning."
      },
      "finish_reason": "stop"
    }
  ]
}

A call on a connection follows the same rules as other Chat calls: text only, the whole answer at once, at most 32 messages and 80,000 UTF-16 code units, and 120 seconds for the model to answer. komplai stores the call's metadata and the model that answered, not the messages. In addition:

reasoning_effort can only be none; any other value is refused with reasoning_not_enabled.
A connection to Anthropic takes a temperature of at most 1, only text as response_format, and at least one message that is not a system message. Anything else is refused with invalid_input.
A model other than the connection's own is refused with model_not_enabled.

The API never moves a call from a connection to komplai's processing or to another connection, also not when the connection fails. The Chat reference has every field. A call on a connection can also stop with these errors:

Code	Status	What happened
connection_not_found	404	No connection with this id is available to you through the API.
processing_choice_not_allowed	403	The connection is not made available through the API.
connection_model_unavailable	409	The provider does not know the model the connection names.
connection_rate_limited	429	The provider refused the call for now. Wait the seconds in `Retry-After`.
connection_unauthorized	502	The provider refused the connection's key.
connection_payment_required	502	The provider wants payment on your organisation's account.
connection_request_refused	502	The provider refused the request, for example a value it does not take.
connection_failed	502	The provider could not be reached or failed.
connection_unavailable	503	The connection's key cannot be read, or its address is refused.
timeout	504	The model did not answer within 120 seconds.