One key, every dialect.

A single inference gateway in front of magnetar-1, built on Qwen 3.6 27B and served on our infrastructure. It implements the OpenAI Chat Completions protocol and the Anthropic Messages protocol in full, including streaming event grammars, tool calling, and native error shapes. Nothing is translated in your client.

Message content is never stored. Logs carry metadata only: timestamps, token counts, latency, status. What you send and what the model returns exist in memory for the duration of the request and nowhere else.

Web chat and API calls draw from the same balance and the same usage pools. There is no separate billing wall between them.

Create a key in settings → API (it is shown once), then pick the dialect you already have an SDK for.

curl https://api.demimonde.dev/v1/chat/completions \
  -H "Authorization: Bearer dm-live-…" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "magnetar-1",
    "messages": [{"role": "user", "content": "…"}],
    "stream": true
  }'

One key space serves every endpoint. Send it as a bearer token or, on Anthropic routes, as the x-api-key header. Either is accepted everywhere.

Authorization: Bearer dm-live-…

# or, equivalently on /v1/messages routes
x-api-key: dm-live-…

Keys are displayed once at creation and stored hashed. You can name them, set a monthly spend cap per key, and revoke them from the dashboard. Last used metadata is visible per key.

GET /v1/me returns your tier, the binding usage window as a percentage with its reset time, and your prepaid balance. The full machine readable spec lives at /.well-known/openapi.yaml.

Set stream: true (OpenAI) or use the messages streaming flow (Anthropic) and the gateway emits the native event grammar for the dialect you chose. There is no intermediate format.

# OpenAI dialect
data: {"type":"chat.completion.chunk","choices":[{"delta":{"content":"…"}}]}

# Anthropic dialect
event: message_start
event: content_block_start
event: content_block_delta   → input_text_delta
event: content_block_stop
event: message_delta         → stop_reason, usage
event: message_stop

Disconnect mid stream and you are billed only for tokens actually emitted, measured from upstream usage. The abort propagates to the GPU worker within half a second.

One model, served on our own infrastructure.

Dated snapshots are immutable, so you can pin a version and never be surprised by a behavior change. Snapshots are deprecated with 90 days of notice. The stable alias always points at the current production model.

Errors arrive in the error shape of the dialect you called, with the correct HTTP status. There is no cross dialect error soup.

// OpenAI dialect
{"error": {"message": "…", "type": "invalid_request_error", "code": "…"}}

// Anthropic dialect
{"type": "error", "error": {"type": "invalid_request_error", "message": "…"}}

Cold starts return a retry_after hint and streaming requests may begin with a warming status event while the worker comes up. Subscribers ride the priority queue.

Rolling usage windows are tracked server side and exposed only as a percentage through /v1/me, never as raw token counts.

Prepaid balance is shared across web chat and API. Deposits in crypto (BTC, XMR, USDT, ETH) earn +10% credit. Subscriptions include usage pools measured against the same rate card, and overage meters seamlessly to your balance instead of hard blocking mid stream.

The gateway was designed for tools like these. Same key, same balance, pick your dialect.

{
  "provider": "openai",
  "baseURL": "https://api.demimonde.dev/v1",
  "apiKey": "dm-live-…",
  "model": "magnetar-1"
}

Provider:       OpenAI Compatible
Base URL:       https://api.demimonde.dev/v1
API Key:        dm-live-…
Model ID:       magnetar-1

# librechat.yaml
endpoints:
  custom:
    name: "Demimonde"
    baseURL: "https://api.demimonde.dev/v1"
    apiKey: "dm-live-…"
    model: "magnetar-1"

Chat Completion Source:  Custom (OpenAI compatible)
Custom Endpoint:         https://api.demimonde.dev/v1
API Key:                 dm-live-…
Model:                   magnetar-1