Book a call

CHOOSE · SOVEREIGNTY

You know where your data goes, and who decides.

The platform is hosted in France or Switzerland, and you choose the models yourself, from a catalogue of approved models.

Your agents

Ascenzia Sentinel
Claude
Mistral
Gemini
GPT (OpenAI)
You switch lines when you choose to; if a model fails, the points switch on their own.

SOVEREIGNTY

Three questions people tend to lump together.

Where your data is processed, who decides what your agents do, and what happens on the day you change model.

[01]

LOCATION

France or Switzerland.

Hosting follows the client's country, and the section just below gives the details.

[02]

REAL CONTROL

The decisions are yours.

You choose the approved models, set the guardrails and fix each agent's budget, and the execution log keeps a record of it.

[03]

FREEDOM OF MODEL

Change models freely.

If a model stops suiting you, you move to another without rewriting your agents, and your memory stays where it is.

Hosting in France tells you nothing about who decides, and we think these three points deserve to be checked one at a time, with us as with any other supplier.

What runs where, and who decides.

In France, in the Paris region, for our French clients. In Switzerland, with our partner Hidora, for our Swiss clients.

  1. 01 · THE PLATFORM

    • The interface, your agents' contextual memory and the database are hosted by us, in France (Paris region) or in Switzerland with Hidora, depending on your country.
  2. 02 · THE MODELS

    • You pick your models from a catalogue. The platform settings show you, model by model, where each request is processed, and you decide which ones to switch on.
  3. 03 · WHEN YOU CHANGE MODEL

    • Your memory and your data stay put. They remain in France (in Switzerland for our Swiss clients); only the model handling the request changes.
Map: hosting in France (Paris region) and in Switzerland

What we see. What we never see.

WE SEEstatuses · durations · tokens

WE NEVER SEEprompts · responses

[01]

METADATA ONLY

Guardrails read each request as it passes through and keep only the verdict in the control logs, plus a little metadata (status, duration, token count). The text of your prompts and the models' responses never appears there.

[02]

YOUR OWN API KEYS, IF YOU PREFER

The catalogue of models and the rules stay in your hands. If you would rather, usage can be charged to your own API keys.

[03]

ISOLATED EXECUTION

An agent runs in an isolated environment and can reach only the tools declared in the connector registry. If your ERP connector is not on that list, the agent simply cannot see it.

Two founders, 25 years of IT experience each, building systems that cannot afford to fail.

We refuse to let any model become a dependency.

Your agents' requests go through the multi-model gateway, and you choose each agent's model from the catalogue. You change it when you decide to, without rewriting your agents or taking out a subscription with every vendor, and your contextual memory stays in France throughout. If a model goes down, the gateway fails over to the fallback you planned.

For each agent, the model you chose from the catalogue and its fallback, ready for the day a model goes down.

Multi-model gatewayModel chosen per agent
AgentModelFallbackStatus
HR questionsClaude SonnetMistral Largeactive
Bank reconciliationMistral LargeClaude Sonnetactive
Meeting summariesGemini FlashMistral Smallfailover
TranslationMistral SmallGemini Flashactive

Demo data

01

ROUTING

The task goes to the approved model that suits it.

02

CHANGING MODEL

You move from one model to another whenever you like, and the gateway fails over to the fallback by itself if one stops responding.

03

CAP

Each agent has a cap in euros, and you get an alert before it is reached.

REGULATORY FRAMEWORK

The platform helps you comply with DORA, NIS2, the AI Act, the Cloud Act and GDPR.

Your security questions

In France or Switzerland, depending on your country, and it stays there when you change model; the hosting section above has the details.

In the control logs, metadata only (the status of each request, how long it took, the number of tokens); the text of your prompts and the responses is not kept there. Your agents' contextual memory is hosted on the platform, in France or Switzerland.

It gives you what you need to document what your agents do, with the approved models, guardrails, each agent's budget and the execution log, which you can export. It helps, and the compliance work itself remains yours (or your DPO's).

There are two separate questions. What we host (the interface, the contextual memory, the database) sits in France or Switzerland. For the models, it depends on which ones you enable, since a model running in Europe does not raise the same issue as one running in the United States; the settings show you where each one runs, and you decide model by model.

Yes, whenever you decide and without rewriting your agents, because the multi-model gateway sits in between. Your memory and your data stay where they are. The gateway also covers you during an outage, by failing over to the fallback model.

Another question? Ask your security questions

FRUGALITY · RESPONSIBLE AI

A model sized to the task.

Paying for a large model to sort emails makes little sense, in our view. The gateway sends simple tasks to a light model and keeps the more powerful ones for work that needs them.

See the figures

Twenty minutes to scope your first agent.

YOUR AI AGENTS. STEERED. SECURED. SOVEREIGN.