# Inference API: point your OpenAI client at the EU

An OpenAI-compatible API served on EU infrastructure. Change one base_url and your existing code runs on EU-resident models, metered per use and recorded from the first token.

## Point your OpenAI client at the EU.

The Akumi Inference API is OpenAI-compatible and served on EU infrastructure. Change one base_url and your existing code runs on EU-resident models: chat, tools, streaming and structured output, metered per use and recorded from the first token.

one base_url · your client · unchanged

The commented line is the only one that changes.

## Change one line. Keep your whole stack.

No new SDK, no rewrite, no migration project. Keep the OpenAI client you already build with, point it at Akumi, and add your key. Chat, function calling, streaming and JSON mode behave exactly as they do today, now on EU-resident models with one metered bill.

## Many models, one endpoint.

The model router puts every model behind the same base_url. Pick an EU-resident model, or allow an external one under your policy. Set routing rules, automatic fallback, and bring your own provider keys.

1. Route by data class: Decide per data class which models are allowed, with crown-jewel data kept EU-only.

2. Survive a provider hiccup: A provider hiccup fails over to the next allowed model, so a request rarely just errors.

3. Bring your own: Use your own provider credentials when you want, billed and audited through one place.

## Your stack runs unchanged.

The endpoints, request bodies and responses match the OpenAI API, so the libraries, frameworks and tools you already use just work.

### Chat Completions

POST /v1/chat/completions with the full message and parameter surface you already send.

### Streaming

Token-by-token server-sent events, the same response you get from stream=true.

### Tools and function calling

Define tools, let the model call them, and steer with tool_choice. Tool content is checked too.

### Structured output

response_format with a JSON schema for reliable, typed responses your code can trust.

### Embeddings

POST /v1/embeddings on EU-hosted models for search, clustering and retrieval.

### Native /v1/process

A native endpoint when you want the full Akumi pipeline response, residency and services included.

## Pay for tokens, not seats.

Base inference is metered per token and request. Every service you switch on is metered on its own line, so the bill reads like the work the platform actually did. No seats, no minimums, no surprise tiers.

### Per token and request

Base inference is metered on real usage and drawn from prepaid credits.

### Per service

Firewall, Recall, Router and Cache each meter independently, on the calls they ran on.

### One transparent bill

A single ledger of credits with per-line detail, exportable for finance and compliance.

## On the record for every call.

### EU-resident inference

Requests are served on EU infrastructure. Inference and storage stay in the EU, with no US CLOUD Act exposure.

### Egress guard

Routing to a non-EU model is blocked unless you explicitly allow it, with personal data pseudonymized before it ever leaves.

### Recorded from the first token

Every request is metered and written to a metadata-only audit trail with its model, region and services. Hand compliance the log.

## Questions people ask about the Inference API.

### Do I have to change my code?

One line. It is an OpenAI-compatible endpoint, so you change the base_url and keep your existing client. Chat, tools, streaming and structured output work as they did.

### Which models can I call?

EU-resident models by default, and external ones where your policy allows them. The model is a string in the request, and GET /v1/models lists what you may send.

### Where does a request actually run?

On EU infrastructure, unless you have allowed a specific model that is not. Every response carries the region that served it.

### How is it billed?

Per token and request, drawn from prepaid credits, with each service you switch on metered on its own line. No seats and no minimums.

### What is stored about my requests?

Metadata only: what ran, when, in which region and under which policy. Prompt content is never stored.

## Start with one base_url.

Create a key, point your client at Akumi, and ship on EU-resident models today.