On-premise AI for board papers, or AI switched off
Which AI model sees board text in Mithaq: one resolver, a model on your own network via an OpenAI-compatible endpoint, a per-board off switch, and a clear leave-or-stay label.
Can Mithaq use an AI model that runs on our own servers?
Yes, for the text features that go through Mithaq's provider resolver: draft minutes from a transcript, agenda drafts and summaries. Point the endpoint at any OpenAI-compatible server on your network, such as vLLM or Ollama, and Mithaq marks the provider as not leaving the environment. A board can also switch AI off entirely. Audio transcription and a few older assistants still use hosted providers, and they stay off when a board keeps AI on its own endpoint or leaves their keys unset.
- One answer to "where does this text go"
- A model on your own rack
- What stays hosted, stated plainly
- How this fits the on-premise edition
One answer to "where does this text go"
Board papers are among the most sensitive documents an organisation holds, and "press Summarise" is a data decision. Mithaq routes its main text features through one resolver, and every screen that offers an AI action can say which provider will see the text and whether it leaves the deployment.
The resolver checks four things in order, and the first match wins. First, the board's own AI policy: if it says AI is off, AI is off for that board whatever keys the platform holds. Second, the operator's setting for the provider type. Third, which keys are present, with a customer endpoint preferred over a hosted one. Fourth, if nothing is configured, AI is off. The resolver itself never makes a network call.
A model on your own rack
Set the endpoint, the model name and a key, and Mithaq sends its requests in the OpenAI chat-completions format to that address. vLLM, Ollama and other servers that speak the same format work without code changes. When the endpoint is on a private address, a local hostname or an internal domain, Mithaq records that the text does not leave the environment and shows that in Settings and in the health report. The operator can also declare the processing region, and Mithaq shows it as declared rather than guessing.
A single board can carry its own endpoint in its policy file, so one organisation can keep a sensitive committee on an internal model while other boards use the platform default.
What stays hosted, stated plainly
Three features go through the resolver today: minutes drafted from a transcript, agenda drafts and summaries. Others do not yet. Recording transcription sends audio to Google Gemini, and it refuses to run for a board whose AI is on its own endpoint, and for any board until an administrator records in the product that the key is on a paid Google project and that the board allows it. The translation assistant, the board chat assistant and the TypeSafe Jev checks call hosted services, and Jev runs only for boards that already use a hosted provider.
On an install with no hosted keys, those features are unavailable and make no call. We list them here because a buyer asking "does any board text leave our network" deserves a feature-by-feature answer, not a slogan.
How this fits the on-premise edition
Mithaq's on-premise edition runs as a single tenant on your servers, installed with Docker Compose. The AI settings above are how that edition handles AI: an internal model, or none. The on-premise page lists what is validated before any offline commitment (network, identity, notifications, backups, updates and integrations), and the pricing page publishes the licence fee.
Prices are published, and the free plan needs no sales call.
See pricing