Providers

Scaleway Paris, Tensorix Dublin, AWS Bedrock `eu-central-1` — what each one hosts.

CleverRouter routes to EU-resident inference providers. Every model runs in a European data centre. No US fallback, no us-east-1, no non-EU shadow routing.

At a glance

Provider	Region	Strengths	Activation
Scaleway	Paris (`fr-par`)	Chat, embed, audio, Pixtral vision	Always on
Tensorix	Dublin (`ie-dub`)	Chat, embed, fast Whisper	If `TENSORIX_API_KEY` set
AWS Bedrock	`eu-central-1`	Cohere embed + rerank, Amazon Nova	If `AWS_BEDROCK_API_KEY` set
Azure AI Foundry	EU Data Boundary (Sweden)	GPT series + Foundry models, embed	Dark-launched — dormant until `AZURE_API_KEY` set

Scaleway (Paris, FR)

OpenAI-compatible passthrough to https://api.scaleway.ai/<project-id>/v1.

Endpoint	Notes
Chat	Streaming, tool calls, vision (Pixtral), JSON mode
Embeddings	`baai/bge-multilingual-gemma2`, `qwen/qwen3-embedding`
Audio	`openai/whisper-large-v3`
Rerank	Not supported

OpenAI-compatible passthrough to https://api.tensorix.ai/v1. The fast-whisper variant beats vanilla Whisper on throughput. Inference runs on Tensorix-owned EU GPUs in Dublin, Ireland (primary) with Helsinki, Finland as failover.

Endpoint	Notes
Chat	Streaming, tool calls
Embeddings	`qwen/qwen3-embedding-8b`
Audio	`systran/faster-whisper-large-v3`
Rerank	Not supported

AWS Bedrock (`eu-central-1`)

Special case: Bearer API key (ABSK format, not SigV4) plus a custom Converse-schema mapping. Active when the gateway has AWS_BEDROCK_API_KEY set.

Endpoint	Notes
Chat	Streaming, tool calls, vision (Anthropic, Nova)
Embeddings	`cohere/embed-v4`
Audio	Not supported
Rerank	`cohere/rerank-v3-5`, `amazon/rerank-v1`

Bedrock model IDs

Bedrock model IDs always use the eu. or global. regional prefix — never us.*. CleverRouter refuses to route to non-EU Bedrock models at request validation.

Azure AI Foundry (EU Data Boundary)

OpenAI-compatible passthrough to the Azure Foundry v1 API (https://<resource>.services.ai.azure.com/openai/v1/, api-key header). One endpoint and one key cover both the GPT series and the Foundry "sold by Azure" models (DeepSeek, Llama, Mistral, Grok, Cohere, Phi). Active only when the gateway has AZURE_API_KEY and AZURE_BASE_URL set — it is dark-launched and currently dormant.

Endpoint	Notes
Chat	Streaming, tool calls; Azure content-filter chunks normalised out
Embeddings	OpenAI-compatible embedding deployments
Audio	Not supported
Rerank	Not supported

EU Data Boundary is EU/EFTA, and ZDR is opt-in

Azure runs inside Microsoft's EU Data Boundary, which covers the EU plus EFTA states Norway and Switzerland (both hold a GDPR adequacy status). Only DataZone* (EU Data Boundary) and EU-regional deployment types are allowed — all Global* types are hard-blocked. Azure also keeps prompts for up to 30 days for abuse monitoring by default, so zero data retention is opt-in here, not the default, until Microsoft approves our Modified Abuse Monitoring request.

EU-only invariant

EU-only routing is an architectural constraint, not a setting:

No US providers. No US fallback. No multi-region failover that crosses the EU border.
Bedrock model IDs validated against the eu.* / global.* prefix.
Azure deployments validated against their deployment type and residency scope — DataZone* (EU Data Boundary, EU/EFTA) and EU-regional only; every Global* type is refused.
All gateway containers run in Scaleway Paris (fr-par).

See Trust → Zero data retention and GDPR & DPA for the full data-flow story including sub-processors.

How a provider is chosen per request

For models hosted by more than one provider, the gateway picks in order:

Explicit pin (X-CleverRouter-Provider) — see Provider pinning.
Priority sort (model_providers.priority ASC, lowest wins).
Health filter — providers marked down in the last health probe are skipped (unless all are down).
Disabled or deprecated mappings are ignored.

Full flow with a diagram: Routing overview.

What's coming

More EU regions on Bedrock — pending AWS rollout of additional Anthropic models in eu-central-1.
Possible OVH / Bento.AI integration when they cover models we don't already have through Scaleway or Tensorix.

Catalog — model IDs and kinds.
Auto-sync — how new models appear.
Routing overview — picking provider per request.