Full power of AI, full control of your data.

• Chat and Fusion

  Ask across models, grounded in your own documents, with citations you can open.

  →

• Agents

  Automations that do the work, ask before anything irreversible, and leave a trail.

  →

• Knowledge

  Your documents, searchable by meaning, with access granted per base and per group.

  →

• Code and Data

  Functions, a document store and hosted sites, without leaving the platform.

  →

• API

  An OpenAI compatible endpoint with a budget per key and a receipt per call.

  →

• Routing

  One name in front of your models, and rules that pick which one answers each call.

  →

1. 01

   Bring your models, or use ours.

   Point a deployment at a provider you already pay for, or pick from the catalogue. The name in the request is your own name for the model, so swapping what is behind it changes nothing in your code.

2. 02

   Give it your knowledge.

   Upload documents into a knowledge base. Answers are grounded in what you uploaded and cite the exact passage, with the page it came from. Access is granted per base, per group and per person.

3. 03

   Let it act, within limits.

   Agents reach the outside world through a broker that mints a capability for one run and nothing more. Anything with consequences can wait for a second person to approve it.

1. Encrypted under your own key

   Your content is sealed with a key belonging to your organization alone, so a copy of any database, taken by anybody, is ciphertext.

2. Deletion that destroys the keys

   Closing an organization removes it from every service and then destroys its keys, which makes anything left behind, backups included, permanently unreadable.

3. Data in the European Union

   Processed and stored in the EU, on European infrastructure, under European law.

4. No training on your data

   Your prompts, documents and files are never used to train a model, ours or anybody else's.

5. Prompts are not archived by default

   The trail keeps metadata, token counts and cost. Keeping the full request body is a decision an organization makes, not the default.

6. One organization cannot see another

   Every token carries the organization it belongs to, and every service derives access from that token rather than from anything the client sends.

7. Least privilege, enforced server side

   Roles map to a fixed catalogue of permissions. The interface hides what you may not do and every backend independently refuses it.

8. Retention you set, with legal hold

   Two clocks: the payload is scrubbed, then the record is deleted. A legal hold suspends both until it is lifted.

Read the security model→

1. Metered per unit

   Tokens per model, web searches, seconds of code execution, function runs, stored data per day, document operations and site runtime.

2. Reserved before the call

   The worst case cost of a request is held against the balance before it goes upstream, so a burst of parallel requests cannot spend the same money twice.

3. A budget per key

   Give an integration its own ceiling. When it is reached the key is refused, and the refusal says which limit stopped it.

4. One bill, in euro

   Everything on one invoice, with the usage behind it visible per day, per model and per key.

How billing works→

• Sovereignty and compliance4 Aug 2026

  The AI Act from 2 August 2026: what changed for a company using AI

  Since 2 August 2026 article 50 of Regulation 2024/1689 applies: an AI system that talks to a person must tell them they are talking to a machine, generative output must be machine readable as artificial, and a deployer must disclose deepfakes and generated text that informs the public on matters of public interest.

  Read next→

• Cost4 Aug 2026

  What AI actually costs a team of twelve: the arithmetic

  For twelve people using an AI assistant at a normal pace, with a 20 GB knowledge base and three hundred automation runs a month, the bill lands around 35 euro a month, which is under 3 euro per person. Loading the documents into the knowledge base costs a further 69 euro, once.

  Read next→

All writing→

Is any of it free?

There is no free plan and no trial, and there are two free models: the cheapest chat model and the cheapest embedding one. The allowance belongs to a person and is counted in calls rather than in euro, so a new chat opens on a model that costs nobody anything and a knowledge base can be filled before anything is bought. Above everybody's own allowance the organization has one monthly ceiling of a person's allowance per seat. When an allowance is spent the call is refused with the moment it comes back, never quietly moved onto a model that charges.

Where is our data processed?

In the European Union. The platform runs on European infrastructure and the databases, object storage and audit trails stay in the EU.

Do you train models on our data?

No. Prompts, documents and generated files are used to answer your requests and for nothing else.

Which models can we use?

The ones your organization enables. The catalogue is per organization and an administrator decides what appears for everybody else. Open weight models such as Llama, Mistral, Qwen and gpt-oss run on European infrastructure and cost the same to call as any other entry in the catalogue, so keeping a workload on open source is a choice about price and independence rather than a compromise on how it is used.

Can we use our own provider account?

Yes. A deployment maps a name you choose to a provider and an upstream model, so your requests keep using your name for it.

What happens when the balance runs out?

The next paid request is refused with a reason rather than run on credit. Spend limits per key and per organization work the same way.

Is there a contract or a minimum?

No subscription is required to start. You top up a balance and use it, and you can set a limit so nothing runs beyond it.