Your developers already use something better than Copilot.
The question is not whether they may – it is whether anyone can show the auditor what left the building.
What it looks like in most places
The tool was chosen because it sat in the agreement and was covered by the data processing agreement. Nobody had to decide anything.
The developers know what is better, and many already use it: on their own accounts, out of their own pocket.
Code, configuration and customer data are sent to models nobody approved, and nobody can account for it afterwards.
This is not a tooling question. It is a governance question, and it belongs to management: which data may be sent to which models, and who can show that the rule is kept?
Three levels we never blur
Where a model runs, and whose law applies to the company running it, are two different questions. We keep them apart in three levels.
Claude through Bedrock in Frankfurt or Stockholm, or GPT through Azure in the EU, in your own cloud account. The frontier models, with the data in Europe.
Data in Europe · the provider under American law
Scaleway, OVHcloud, IONOS, STACKIT or T-Systems running DeepSeek, Qwen, GLM, gpt-oss or Mistral. The weights may be Chinese, American or French in origin; they run in Europe at a European company.
Data in Europe · European law
An open model on machines you control. The answer where data may not leave the building at all.
Nothing leaves the building
Level 1 is data in Europe, not European law. We do not call it sovereign, and we say which level each task runs on.
Two tiers: open by default, frontier where the policy allows
The frontier tier
Where the task needs it and the policy allows it: Claude through Bedrock in Frankfurt or Stockholm, or GPT through Azure in the EU. In your own cloud account, on your own agreement.
The open tier
Our starting point is DeepSeek V4.1 Flash at Scaleway. The model is Chinese in origin and runs in Paris; we say both. For those who will not have Chinese weights wherever they run, Devstral at Mistral is the French alternative.
The choice is made per task, not once for the whole company. That is what the gateway is for.
The gateway: one endpoint, a policy and an audit trail
What you buy is the gateway. It stands on your European cloud, is written in Terraform in your own repository, and every tool talks to it.
Every tool talks to one endpoint. The policy decides where the request runs.
The developer’s tools
- OpenCode, Cline, Continuemodel-agnostic, on the developer’s machine
- Claude Codepointed at the frontier tier
- CI and agentsone key per pipeline
The gateway
on your European cloud · Terraform in your repository
- Sign-in with your own users
- Routing policy
- Secret scanning
- Audit log
- Cost per team
Where the model runs
- Level 1 · American provider, European regionClaude via Bedrock, GPT via Azure in the EU · when the policy allows
- Level 2 · European providerDeepSeek V4.1 Flash at Scaleway, Devstral at Mistral · the default
- Level 3 · own infrastructurean open model on your own machines · when data may not leave the building
The evidence pack The log and the policy are what you show the auditor and the customer.
OpenAI-compatible, so the tools the developers already know can use it without rebuilding.
Single sign-on through Keycloak and keys per team and per pipeline, so the log names people and services, not a shared token.
By repository, task or data class. The default is the open model at the European provider; frontier only where the policy allows; own infrastructure where data may not leave the building.
Secret scanning and redaction before anything leaves the gateway.
Who sent what to which model, kept for as long as you decide.
Per team and per model, so the bill for open against frontier is a number and not an estimate.
The package
AI tooling on European soil
- Two to three weeks
- The gateway in your account and your repository
- Two tools wired to it
- The policy and the evidence pack
Already on a European cloud?
Then the road is shorter: the models run where the systems already run, and the gateway goes into the landing zone you have.
The evidence pack
The log and the policy set against the requirements your customers and auditor ask about: NIS2’s requirement to keep the supply chain under control, and the AI Act’s requirement of AI literacy for those who use the tools. With them, the sub-processors and data processing agreements, listed per level.
What keeps it current
The models change every month. Our monthly review of which open models run at which providers in Europe is your letter, and the routing policy is reviewed with it.
We use it ourselves
This is how we work ourselves: OpenCode with DeepSeek V4.1 Flash at Scaleway, and Claude Code with frontier models where the data allows. Nomadics’ own site is built with Claude Code; it is public code, and so it sits on that side of the line.
In November 2026 we run a week of our own work on the open model and publish the hours and the token bill afterwards, beside the same tasks on a frontier model.
Book the package
Write to Martin with how many developers you are and which tools you use today. He answers within one working day with the plan for the two to three weeks and the fixed price.