Your key, your model

You bring your own provider

Under Instellingen → AI you pick Anthropic, OpenAI or "OpenAI-compatibel (eigen server)" — anything speaking the OpenAI chat API on a base URL of yours, from Azure OpenAI to a model you host on your own machine. You paste your own API key into the field; it is stored encrypted and never played back, and on a later save you leave it blank to keep the stored one. The model field is free text with a list beside it, and "Modellen ophalen" fetches your provider's current models, so a new model needs no release from us. "Verbinding testen" sends a real, tiny request through the stored settings and shows your provider's own error verbatim, because that is the text you can actually act on.

• Anthropic, OpenAI or an OpenAI-compatible server on a base URL of your own
• The key is kept encrypted and is never shown again
• Fetching models reads the provider's live list, not a list of ours that rots
• You pay your provider directly; nothing of ours sits between your client data and the model

Eight features, independently switchable

Every AI feature runs through the same core, so one key, one set of toggles and one budget apply everywhere at once. An agency that wants writing help but no chat assistant switches on exactly that and nothing else. Each feature can also carry its own model: the default model is sized for writing, while quick-add on Uren (Time) is better served by a fast one. The Huisstijl field (house style, up to 4000 characters) is where you describe how your agency writes, and that text is prepended to every writing prompt. Features that read what somebody else wrote, such as an e-mail or a meeting, deliberately get a switch of their own: agreeing to have AI polish a paragraph is not agreeing to have it read a client's mailbox.

• Assistent: a slide-over from the top bar that can look up anything in your workspace and records a fixed list of things for you
• Schrijfhulp: inside every rich-text field, with improve, shorten, expand, spelling & grammar, business or informal tone, translate and write out; then replace or insert
• Urenhulp: one line in Snel toevoegen on Uren becomes a draft time entry, plus "Vul mijn dag aan" for the gap between your schedule and what you logged
• Briefing & rapporten: a digest of a client's recent activity, and a monthly draft you edit yourself before it becomes anything
• Taak inspreken, aanpassen en plannen: dictate a task in one breath, change a task in words, and have its steps written from the description
• E-mail in taak: a task you create while approving an e-mail, filled in from that e-mail, written from your agency's side
• E-mail naar taakadres: filling in what a mailed-in task left open, without overwriting what the sender wrote
• Vergadertranscripten en verslagen: transcribing a recorded meeting and drafting its minutes with action items

A budget you watch fill up

You fill in one number under Maandelijks tokenbudget (monthly token budget); empty means unlimited. The Verbruik (usage) panel on the same screen shows requests and tokens per feature for this calendar month, with a bar that turns amber at 80% and red at 100%. For someone at work the cap is soft: they are told the monthly budget is reached and can deliberately continue anyway, while calls with no person behind them stop hard. What gets written down is counts and labels, never the text of a prompt or an answer.

The assistant reads what you may, and writes a fixed list

The sparkle in the top bar opens the Assistent beside your work. On a client, project or task page it inherits that record, so "this client" simply resolves. It searches the same API an external AI client gets over MCP, so it knows about domains, invoices and marketing figures as well as clients, tasks and hours, and it sees exactly what you may see and nothing more. It writes only from a fixed list: create or change a task, post a comment, log hours, and start or stop the timer. Every write goes, as you, through the same checks and onto the same trail as on the screen; configuring something, sending an invoice or publishing a report is beyond it. Every answer carries source chips that link back to the records it came from, so you can check a number instead of believing it.

Speak it instead of typing it

There is a microphone at Snel toevoegen (quick add) on Uren, at Inspreken on Taken, in a task's Aanpassen met schakl box and in the Assistent's input. Speech needs its own credential under Spraak naar tekst (OpenAI, a compatible server or Mistral), because Anthropic has no transcription endpoint at all and is therefore deliberately absent from that list. The audio goes to the service you configure there and nowhere else — the browser's own speech API ships everything to Google in Chrome and Edge, and to Apple in Safari, and that was not a choice to make on your behalf. Whatever you dictate always appears as text first, because a Dutch client name is exactly what a recogniser gets wrong.

• Its own speech provider and its own monthly budget, measured in seconds of audio
• Two minutes for a time entry, five for a task or a question to the assistant, with the cap shown beside the counter
• The transcript lands in the field or in a draft, never straight into a record
• The microphone is only drawn where it can actually work

Off means invisible, and what the licence does

Until a provider is configured, no AI control is drawn anywhere: the product does not nag for a key. At the bottom of Instellingen → AI, "AI-configuratie verwijderen" switches it all back off and keeps the usage history. Who may configure the provider and who may use the features are two separate permissions, granted per role under Instellingen → Rollen (Settings → Roles); configuring is admin-only by default. The AI features are an extension module and therefore need a licence key, while the core of the platform is open source (AGPL). A fresh install gets fourteen days of grace first, and when a licence lapses the settings screen and the usage history stay readable: only generating stops.