Your AI stack, under your own jurisdiction
Most sovereignty claims are promises about vendor behaviour: where data will be stored, who will be allowed to access it, which jurisdiction will apply. Wäg makes a different kind of claim. It has no control plane of its own — so there is no data path anyone could compel, in any jurisdiction, including ours. That is not a policy we maintain; it is a property of the software, and you can check it at your own network layer.
Sovereignty went from talking point to policy
In November 2025 all 27 EU member states signed the Declaration for European Digital Sovereignty; France and Germany convened a dedicated summit the same month. In June 2026 the European Commission's Tech Sovereignty Package introduced the Cloud and AI Development Act — mandating sovereignty risk assessments and moving to bar US hyperscalers from the most sensitive public-sector contracts.
Two questions decide who controls your AI traffic
Whose infrastructure runs the data plane, and whose runs the control plane? Most answers to the second question are not yours. A self-hosted data plane reporting to a vendor-operated control plane still places a fourth party above every request — and US-incorporated suppliers are additionally subject to FISA 702, which permits collection of non-US persons' communications without individualised warrants and without meaningful redress. That authority, not the CLOUD Act, is what the Court of Justice identified when it invalidated Privacy Shield.
Ask any vendor where its control plane runs. The answer is architectural, and it does not change with a region setting.
No control plane of ours, above none of your data
Wäg ships as self-hosted open source with no control plane of its own — no vendor tenancy holding your traffic, no telemetry channel, no licence check phoning home, no managed tier. There is no data path anyone could compel, in any jurisdiction, including ours. That is not a policy we maintain; it is a property of the software, and you can confirm it by watching your own egress.
Your keys, your region, your logs — your control.
The same authority reaches software suppliers
In June 2026 a US export-control directive required Anthropic to suspend access to two frontier models, Fable 5 and Mythos 5, for every foreign national worldwide — including its own employees. Access was restored about two weeks later. For teams calling one provider directly from application code that was a two-week outage; for teams behind a gateway with ordered fallbacks it was a routing event.
That authority is not limited to model providers. US export-control and sanctions law covers software and technology, and it reaches any US-incorporated supplier — including the gateway sitting in front of your entire AI stack. The mechanism is well established: an Entity List designation requires a licence to export US items to the named party, which in 2019 severed an existing software relationship and cost the designated company access to services and updates it had been relying on. The same year, a major US developer platform restricted accounts for users in sanctioned regions. Nothing about that machinery treats infrastructure software differently from a model API.
For an open-source dependency the failure mode is rarely that the binary stops running. Apache-2.0 means you keep the code — it does not mean you keep a maintained dependency. What a directive against your supplier costs you is security patches and CVE response, upgrades, the roadmap, commercial support and any SLA, and any managed tier you were paying for. For the one component in front of every AI request, an unmaintained fork you now own is not a viable steady state; it is an incident with a long tail.
Two things reduce that exposure, and they are independent. Multi-provider federation makes a provider becoming unavailable a configuration change rather than an outage — implemented once at the control plane instead of in every service. And a supplier outside US jurisdiction cannot be issued that class of directive in the first place: Bitkaio Ltd is registered in the United Kingdom, outside FISA 702 and outside US export-control authority.
The Fable 5 and Mythos 5 suspension, June 2026Six things a sovereign gateway has to get right
A European supplier
Wäg is built by Bitkaio Ltd, registered in the United Kingdom — European, outside the EU. That places it outside FISA 702, which permits collection of non-US persons' communications from US providers without individualised warrants or meaningful redress, and outside US export-control directives. EEA personal data reaches the UK under an adequacy decision renewed in December 2025 through to 2031.
Self-hosted, in your perimeter
Wäg runs entirely inside your own infrastructure — your cloud, your region, or fully air-gapped — and residency policies keep each request on model endpoints in the jurisdictions you allow. There is no managed control plane, and no data path back to us to subpoena.
No phone-home
Prompts, completions, keys, logs, and telemetry never leave your environment. Nothing is sent to a vendor endpoint for licensing, analytics, or "improvement" — by design, not by configuration.
Open source, verifiable
Apache-2.0 at the core. Every image is signed with cosign and ships with an SBOM, and the licence is checked offline, so what you run can be verified without asking us — aligned with the "open-source-first" principle at the heart of the EuroStack vision.
Your keys, your vault
Provider credentials resolve from your own HashiCorp Vault, Azure Key Vault, AWS Secrets Manager, or GCP Secret Manager. With vault references, raw keys never reach the database; a key entered inline is stored sealed with AES-256-GCM.
You are the operator
The Cloud and AI Development Act, tabled in June 2026, proposes four sovereignty levels — from EU-located infrastructure through demonstrated third-country independence, EU ownership and control, to full supply-chain transparency. It is a proposal, and the thresholds are not yet fixed. What matters architecturally is that self-hosted software makes your organization the operator: the infrastructure, the personnel and the controls being assessed are yours, not your supplier's.
Change one base URL, not your application
Wäg speaks the same OpenAI-compatible language and imports your existing LiteLLM configuration. LiteLLM is a very good gateway; if what you now need is governance rather than routing, that is a different layer — and adopting it should cost less than living without it.
- 01
Keep your OpenAI or LiteLLM SDK
Wäg is drop-in OpenAI-compatible and a drop-in replacement for the LiteLLM proxy. Point your existing OpenAI or LiteLLM clients at the gateway's base URL — your application code doesn't change.
- 02
Import your LiteLLM config
Already running LiteLLM? waeg import litellm maps your model list, routing strategy, projects and keys. What it cannot map — guardrails, cache settings, callbacks — is listed in the import report to re-create, never silently approximated.
- 03
Gain the AI-native core
Same mental model, but now with an inference-request IR, intelligent routing with capability floors, four-tier semantic caching, and hierarchical budgets — in Rust.
- 04
Run it where you choose
Docker Compose, Helm, or a single embedded binary — in your EU region, on-prem, or air-gapped. No cloud leash, no phone-home, and no vendor control plane above your data plane.
# Bring your existing LiteLLM config across
waeg import litellm --from litellm.yaml -o waeg.yaml
# Point your OpenAI SDK at Wäg — in your EU region
export OPENAI_BASE_URL=https://waeg.internal/v1
# Same call. Your infrastructure. No phone-home.
curl $OPENAI_BASE_URL/chat/completions \
-H "Authorization: Bearer $WAEG_KEY" \
-d '{"model":"chat-auto","messages":[...]}'Run it yourself. Keep control.
Open source · self-hosted · built in Europe · no phone-home · Apache-2.0.